Uploaded by Sarthak Kulkarni

Nahmias

advertisement
Slope = – Q
Inventory (I(t))
Inventory Levels for the EOQ Model
T
Time t
All-units discount average annual cost function
G(Q)
240
230
G0(Q)
220
G1(Q)
210
G2(Q)
200
190
180
100 200 300 400 500 600 700 800 900 1,000 1,100 1,200
Q
Inventory
Changes in Inventory Over Time for Continuous-review (Q, R) System
R
Q
Order
placed
Order Time
arrives
Nahmias-Olsen FrontMatter.fm Page 1 Tuesday, December 30, 2014 10:01 AM
SEVENTH EDITION
P R O D U C T I O N AND
O p e r a t i o n s An a l y s i s
….
Nahmias-Olsen FrontMatter.fm Page 2 Tuesday, December 30, 2014 10:01 AM
SEVENTH EDITION
P R O D U C T I O N AND
O p e r a t i o ns Ana l y s i s
S T R A T E G Y
•
Q U A L I T Y
•
A N A L Y T I C S
Steven Nahmias
Santa Clara University
Tava Lennon Olsen
University of Auckland
WAVELAND
PRESS, INC.
Long Grove, Illinois
•
A P P L I C A T I O N
Nahmias-Olsen FrontMatter.fm Page 3 Tuesday, December 30, 2014 10:01 AM
For information about this book, contact:
Waveland Press, Inc.
4180 IL Route 83, Suite 101
Long Grove, IL 60047-9580
(847) 634-0081
info@waveland.com
www.waveland.com
Copyright © 2015 by Steven Nahmias and Tava Lennon Olsen
10-digit ISBN 1-4786-2306-3
13-digit ISBN 978-1-4786-2306-9
All rights reserved. No part of this book may be reproduced, stored in a retrieval system, or transmitted in any form or by any means without permission in writing from
the publisher.
Printed in the United States of America
7 6
5 4
3 2
1
I dedicate this seventh edition to my lucky seven,
my beautiful wife Vivian, whose birthday falls on
the seventh of February.
Steven Nahmias
To my family.
Tava Lennon Olsen
nah77856_fm_new.indd 5
18/12/14 10:21 PM
Brief Contents
About the Author xv
Preface xvi
Introduction xvii
1
Strategy and Competition 1
2
Forecasting 52
3
Sales and Operations Planning 128
Supplement 1 Linear Programming 169
7 Service Operations Management 367
Supplement 2 Queueing Techniques 417
8
9
10
11
12
13
Push and Pull Production Control
­Systems: MRP and JIT 437
Operations Scheduling 490
Project Scheduling 543
Facilities Layout and Location 593
Quality and Assurance 666
Reliability and Maintainability 742
4
Inventory Control Subject to Known
­Demand 198
5
Inventory Control Subject to Uncertain
­Demand 249
Appendix: Tables 796
6
Supply Chain Management 315
Index 814
vi
nah77856_fm_new.indd 6
18/12/14 10:21 PM
Contents
About the Author xv
Preface xvi
1.11 Capacity Growth Planning: A Long-Term
Strategic Problem 38
Economies of Scale and Economies of Scope 38
Make or Buy: A Prototype Capacity Expansion
Problem 39
Dynamic Capacity Expansion Policy 40
Issues in Plant Location 44
Problems for Section 1.11 46
Introduction xvii
Chapter 1
Strategy and Competition 1
Chapter Overview 1
1.1
Snapshot Application: Apple Adopts a New
­Business Strategy and Shifts Its Core
­Competency from Computers to Portable
Electronics 3
Manufacturing Matters 5
Manufacturing Jobs Outlook 7
1.2
A Framework for Operations Strategy 8
1.3
Competing in the Global Marketplace 10
Strategic Dimensions 9
Snapshot Application: Global Manufacturing
Strategies in the Automobile Industry 13
Problems for Sections 1.1–1.3 13
Strategic Initiatives: Reengineering the
­Business Process 14
1.5 Strategic Initiatives: Just-in-Time 17
1.6Strategic Initiatives: Time-Based
Competition 19
1.7 Strategic Initiatives: Competing on
Quality 20
1.4
1.8
1.9
1.12 Summary 47
Additional Problems for Chapter 1 48
Appendix 1–A Present Worth Calculations 50
Bibliography 51
Chapter 2
Forecasting 52
Chapter Overview 52
2.1 The Time Horizon in Forecasting 55
2.2 Characteristics of Forecasts 56
2.3 Subjective Forecasting Methods 57
2.4 Objective Forecasting Methods 58
2.5
2.6
Problems for Sections 1.4–1.7 22
Strategic Initiatives: Servicization 23
Servicization: Moving Downstream 23
The IBM Story 24
Performance Based Contracts 25
Leasing Versus Buying 25
Green Leasing 26
Problems for Section 1.8 27
2.7
Matching Process and Product Life Cycles 27
The Product Life Cycle 27
The Process Life Cycle 28
The Product–Process Matrix 29
Problems for Section 1.9 31
1.10 Learning and Experience Curves 31
Learning Curves 32
Experience Curves 34
Learning and Experience Curves and
­Manufacturing Strategy 36
Problems for Section 1.10 36
2.8
Causal Models 58
Time Series Methods 59
Snapshot Application: Advanced Forecasting,
Inc., Serves The Semiconductor Industry 60
Problems for Sections 2.1–2.4 60
Notation Conventions 62
Evaluating Forecasts 62
Problems for Section 2.6 64
Snapshot Application: Pfizer Bets Big on
­Forecasts of Drug Sales 65
Methods for Forecasting Stationary Series 66
Moving Averages 66
Problems on Moving Averages 69
Exponential Smoothing 69
Multiple-Step-Ahead Forecasts 73
Comparison of Exponential Smoothing and
­Moving Averages 74
Problems for Section 2.7 75
Snapshot Application: Sport obermeyer slashes
costs with improved forecasting 76
Trend-Based Methods 77
Regression Analysis 77
Problems for Section 2.8 78
Double Exponential Smoothing Using Holt’s
Method 79
More Problems for Section 2.8 80
vii
nah77856_fm_new.indd 7
18/12/14 10:21 PM
viii Contents
2.9
Methods for Seasonal Series 81
Seasonal Factors for Stationary Series 81
Determining the Deseasonalized Series 83
Problems for Section 2.9 84
Snapshot Application: Nate Silver Perfectly
Forecasts 2012 Presidential Election 85
Winters’s Method for Seasonal Problems 85
More Problems for Section 2.9 90
2.10 Box-Jenkins Models 90
Estimating the Autocorrelation Function 91
The Autoregressive Process 94
The Moving-Average Process 95
Mixtures: ARMA Models 97
ARIMA Models 97
Using ARIMA Models for Forecasting 99
Summary of the Steps Required for Building
ARIMA Models 100
Case Study. Using Box-Jenkins Methodology
to Predict Monthly International Airline
­Passenger Totals 101
Snapshot Application: A Simple Arima Model
Predicts The Performance of The U.S.
Economy 105
Box-Jenkins Modeling—A Critique 105
Problems for Section 2.10 105
Chapter 3
Sales and Operations Planning 128
Chapter Overview 128
3.1 The S&OP Process 130
3.2
Key Performance Indicators 135
3.3
The Role of Uncertainty 137
3.4
Aggregate Planning of Capacity 138
3.5
2.11 Practical Considerations 106
3.6
2.12 Overview of Advanced Topics in
­Forecasting 108
3.7
Model Identification and Monitoring 106
Simple versus Complex Time Series
­Methods 107
Simulation as a Forecasting Tool 108
Forecasting Demand in the Presence of
Lost Sales 107
2.13 Linking Forecasting and Inventory
­Management 111
Case Study. Predicting the Growth of
Facebook 112
2.14 Historical Notes and Additional Topics 116
2.15 Summary 117
Additional Problems on Forecasting 117
Appendix 2–A Forecast Errors for
Moving Averages and Exponential
Smoothing 122
Appendix 2–B Derivation of the Equations
for the Slope and Intercept for Regression
Analysis 124
Appendix 2–C Glossary of Notation
for Chapter 2 126
Bibliography 126
nah77856_fm_new.indd 8
Snapshot Application 133
Problems for Section 3.1 134
3.8
3.9
Problems for Section 3.2 136
Problems for Section 3.3 138
Costs in Aggregate Capacity Planning 139
A Prototype Problem 141
Chase, Level, and Mixed Strategies 143
Problems for Sections 3.4 143
Snapshot Application: HP Enterprise Services
Uses Optimization for Workforce ­Planning 144
Solving Aggregate Planning Problems 145
Cost Parameters and Given Information 145
Problem Variables 146
Problem Constraints 146
Rounding the Variables 147
Extensions 148
Problems for Section 3.5 154
Disaggregating Plans 157
Aggregate 157
Snapshot Application: Welch’s Uses A
­ ggregate
Planning for Production S­ cheduling 159
Problems for Section 3.6 160
Sales and operation Planning on a Global
Scale 160
Historical Notes 161
Summary 162
Additional Problems on Aggregate P
­ lanning 163
Appendix 3–A Glossary of Notation
for Chapter 3 167
Bibliography 167
Supplement 1 Linear Programming 169
S1.1 Introduction 169
S1.2 A Prototype Linear Programming Problem 169
S1.3 Statement of the General Problem 171
Definitions of Commonly Used Terms 172
Features of Linear Programs 173
S1.4 Solving Linear Programming Problems
Graphically 174
Graphing Linear Inequalities 174
Graphing the Feasible Region 176
Finding the Optimal Solution 177
18/12/14 10:21 PM
Contents ix
Problems for Section 4.7 226
Snapshot Application: SmartOps Assists in
Designing Caterpillar’s Inventory Control
System 227
Identifying the Optimal Solution Directly
by Graphical Means 179
S1.5 The Simplex Method: An Overview 180
S1.6 Solving Linear Programming Problems
with Excel 181
Entering Large Problems Efficiently 185
S1.7 Interpreting the Sensitivity Report 187
Shadow Prices 187
Objective Function Coefficients and
Right-Hand Sides 188
Adding a New Variable 188
Using Sensitivity Analysis 189
S1.8 Recognizing Special Problems 191
Unbounded Solutions 191
Empty Feasible Region 192
Degeneracy 194
Multiple Optimal Solutions 194
Redundant Constraints 194
S1.9 The Application of Linear Programming to
Production and Operations Analysis 195
Bibliography 197
Chapter 4
Inventory Control Subject to Known
­Demand 198
Chapter Overview 198
4.1 Types of Inventories 201
4.2 Motivation for Holding Inventories 202
4.3 Characteristics of Inventory Systems 203
4.4 Relevant Costs 204
4.5
4.6
4.7
Problems for Section 4.8 231
4.9 EOQ Models for Production Planning 231
Problems for Section 4.9 235
4.10 Power-of-Two Policies 236
Case Study. Betty Buys a Business 238
4.11 Historical Notes and Additional Topics 240
4.12 Summary 241
Additional Problems on Deterministic
Inventory Models 242
Appendix 4–A Mathematical Derivations for
Multiproduct Constrained EOQ Systems 246
Appendix 4–B Glossary of Notation for
Chapter 4 248
Bibliography 248
Chapter 5
Inventory Control Subject to Uncertain
­Demand 249
Chapter Overview 249
The Nature of Randomness 254
Optimization Criterion 256
5.3
The Newsvendor Model 258
The EOQ Model 210
The Basic Model 210
Inclusion of Order Lead Time 213
Sensitivity 214
EOQ and JIT 215
Problems for Section 4.5 216
Extension to a Finite Production Rate 218
Overview of Models Treated in this Chapter 253
5.1
5.2
Holding Cost 204
Order Cost 206
Penalty Cost 207
Problems for Sections 4.1–4.4 208
5.4
Problems for Section 4.6 219
Quantity Discount Models 220
Optimal Policy for All-Units Discount Schedule 221
Summary of the Solution Technique for
All-Units Discounts 223
Incremental Quantity Discounts 223
Summary of the Solution Technique
for Incremental Discounts 225
Other Discount Schedules 225
nah77856_fm_new.indd 9
*4.8 Resource-Constrained Multiple Product
­Systems 227
5.5
Problems for Sections 5.1 and 5.2 257
Notation 258
Development of the Cost Function 259
Determining the Optimal Policy 260
Optimal Policy for Discrete Demand 262
Extension to Include Starting Inventory 262
Snapshot Application: Using Inventory Models
to Manage the Seed-Corn Supply Chain at
Syngenta 263
Extension to Multiple Planning Periods 264
Problems for Section 5.3 265
Lot Size–Reorder Point Systems 267
Describing Demand 268
Decision Variables 268
Derivation of the Expected Cost Function 268
The Cost Function 270
Inventory Level versus Inventory Position 272
Snapshot Application: Inventory Management
Software for the Small Business 273
Service Levels In (Q, R) Systems 274
Type 1 Service 274
Type 2 Service 274
18/12/14 10:21 PM
x Contents
5.6
5.7
Optimal (Q, R) Policies Subject to Type 2
­Constraint 275
Imputed Shortage Cost 277
Scaling of Lead Time Demand 277
Estimating Sigma When Inventory Control
and Forecasting Are Linked 278
*Lead Time Variability 279
Calculations in Excel 279
Negative Safety Stock 280
Problems for Sections 5.4 and 5.5 280
Additional Discussion of Periodic-Review
Systems 282
(s, S) Policies 282
*Service Levels in Periodic-Review S­ ystems 283
Fixed Order Size Model 284
Problems for Section 5.6 284
Snapshot Application: Tropicana Uses Sophisticated
Modeling for Inventory Management 285
Multiproduct Systems 286
ABC Analysis 286
Exchange Curves 288
Problems for Section 5.7 290
6.2
6.3
6.4
6.5
6.6
*5.8 Overview of Advanced Topics 291
Multi-echelon Systems 291
Perishable Inventory Problems 293
Snapshot Application: Intel Uses Multiechelon
Inventory Modelling to Manage the Supply
Chain for Boxed CPUs 294
6.7
Historical Notes and Additional
Readings 295
5.10 Summary 296
6.8
Appendix 5–A Notational Conventions and
Probability Review 303
Appendix 5–B Additional Results and
Extensions for the Newsvendor Model 304
Appendix 5–C Derivation of the Optimal
(Q, R) Policy 308
Appendix 5–D Probability Distributions for
Inventory Management 308
Appendix 5–E Glossary of Notation for
Chapter 5 312
6.9
5.9
Additional Problems on Stochastic
Inventory Models 297
Bibliography 313
Chapter 6
Supply Chain Management 315
Chapter Overview 315
6.1
The Supply Chain as a Strategic Weapon 318
Supply Chain Strategy 319
Snapshot Application: Wal-Mart Wins with Solid
Supply Chain Management 320
nah77856_fm_new.indd 10
The Role of Information in the Supply
Chain 321
Snapshot Application: Anheuser-Busch
Re-engineers their Supply Chain 322
The Bullwhip Effect 322
Electronic Commerce 326
RFID Technology 327
Problems for Sections 6.1 and 6.2 329
The Transportation Problem 329
Generalizations of the Transportation
­Problem 333
Infeasible Routes 334
Unbalanced Problems 334
More General Network Formulations 335
Snapshot Application: IBM Streamlines Its Semiconductor Supply Chain Using Sophisticated
Mathematical 338
Problems for Sections 6.3–6.5 338
Determining Delivery Routes in Supply
Chains 341
Practical Issues in Vehicle Scheduling 345
Snapshot Application: J.B. Hunt Saves Big with
Routing and Scheduling A
­ lgorithm 346
Problems for Section 6.6 346
Risk Pooling 347
Inventory/Location Pooling 348
Product Pooling and Postponement 351
Capacity Pooling 352
Designing Products for Supply Chain
­Efficiency 353
Additional Issues in Supply Chain Design 354
Snapshot Application: Dell Computer Designs
the Ultimate Supply Chain 356
Multilevel Distribution Systems 356
Problems for Sections 6.7–6.9 358
6.10 Incentives in the Supply Chain 359
Problems for Section 6.10 362
6.11 Global Supply Chain management 362
Problems for Section 6.11 363
6.12 Summary 364
Bibliography 365
Chapter 7
Service Operations Management 367
Chapter Overview 367
7.1 Service Operations Strate Gy 370
The Service Economy 370
Service Quality 371
Measuring Quality 372
Controlling Quality 372
Paying for Quality 372
18/12/14 10:21 PM
Contents xi
7.2
7.3
7.4
7.5
7.6
7.7
7.8
Key Decisions for Service Businesses 373
Snapshot Application: Southwest Airlines
­Competes with Service 374
Managing Variability 374
Service Competition 375
Problems for Section 7.1 376
Flow Systems 377
Process Flow Diagrams 377
Capacity 378
Flow Rates and Utilization 378
Problems for Section 7.2 379
Modeling Unscheduled Arrivals 380
Poisson Process 380
Exponential Interarrival Times 382
General Arrival Processes 383
Pooling in Services 384
Probability of Delay 384
Problems for Section 7.3 385
Queueing Systems 386
Structural Aspects of Queueing Models 387
Notation 387
Little’s law 388
The M/M/1 Queue 389
Problems for Section 7.4 391
General Queueing Models 392
Expected Time in System for a Single Server
System 393
Multiple Parallel Servers 395
Systems with Abandonment 396
Priorities 397
Other Queueing Extensions 399
Simulation 399
Improving a Service Process 400
Problems for Section 7.5 401
The Human Element in Service
­Systems 401
Snapshot Application: Using Queueing to
make Staffing Decisions Saves Lives 402
The Psychology of Queueing 402
Introducing Technology into Services 403
Guidelines for Service Guarantees and
­Refunds 404
Snapshot application: Disney uses both the
­Science and the Psychology of Queueing 405
Revenue Management 407
Airline Revenue Management Overview 407
Revenue Management Basics 408
nah77856_fm_new.indd 11
Historical Notes and Additional
­Readings 411
7.10 Summary 411
7.9
Additional Problems for Chapter 7 412
Appendix 7–A Simulation Implementation 414
Random Number Generation 414
Entity Driven Logic 414
Bibliography 415
Supplement 2 Queueing Techniques 417
S2.1 Details of the Poisson Process and Exponential
Distribution 417
S2.2 Analysis of the M/M/1 Queue 418
Waiting Time Distribution 421
S2.3 Further Results for M/M Queues 422
The M/M/s Queue 424
The M/M/1 Queue with a Finite Capacity 427
S2.4 Infinite Server Results 430
The M/g/∞ queue 430
Infinite Server Limits 431
S2.5 Queueing Networks 432
S2.6 Optimization of Queueing Systems 433
Typical Service System Design Problems 433
Modeling Framework 434
Bibliography 436
Chapter 8
Push and Pull Production Control Systems:
MRP and JIT 437
Chapter Overview 437
MRP Basics 440
JIT Basics 442
8.1
The Explosion Calculus 443
8.2
Alternative Lot-Sizing Schemes 449
Call and Contact Centers 405
Call Center Basics 405
Metrics 406
Call Routing 407
Lead Time Pricing 409
Nontraditional Applications for Revenue
­Management 410
Problems for Sections 7.6–7.8 410
Problems for Section 8.1 447
EOQ Lot Sizing 449
The Silver–Meal Heuristic 450
Least Unit Cost 451
Part Period Balancing 452
Problems for Section 8 .2 453
8.3
Incorporating Lot-Sizing Algorithms into the
Explosion Calculus 455
8.4
Lot Sizing with Capacity C
­ onstraints 457
Problems for Section 8 .3 456
Problems for Section 8 .4 460
18/12/14 10:21 PM
xii Contents
8.5
8.6
8.7
8.8
8.9
8.10
Shortcomings of MRP 461
Uncertainty 461
Capacity Planning 462
Rolling Horizons and System Nervousness 463
Additional Considerations 465
Snapshot Application: Raymond Corporation Builds
World-Class Manufacturing with MRP II 466
Problems for Section 8 .5 467
JIT Fundamentals 468
The Mechanics of Kanban 468
Single Minute Exchange of Dies 470
Advantages and Disadvantages of the
Just-in-Time Philosophy 471
Implementation of JIT in the United States 474
Problems for Section 8 .6 475
A Comparison of MRP and JIT 476
JIT or Lean Production? 477
Historical Notes 478
Summary 479
Additional Problems for Chapter 8 480
Appendix 8–A Optimal Lot Sizing for
­ Time-Varying Demand 484
Appendix 8–B Glossary of Notation for
Chapter 8 488
Bibliography 489
Chapter 9
Operations Scheduling 490
Chapter Overview 490
9.1 Production Scheduling and the Hierarchy
of Production Decisions 493
9.2 Important Characteristics of Job Shop
­Scheduling Problems 495
9.3
9.4
9.7
Objectives of Job Shop Management 495
Job Shop Scheduling Terminology 496
A Comparison of Specific Sequencing
Rules 498
First-Come, First-Served 498
Shortest Processing Time 499
Earliest Due Date 499
Critical Ratio Scheduling 500
9.5
Objectives in Job Shop Management: An
­Example 501
9.6
An Introduction to Sequencing Theory for a
Single Machine 503
Problems for Sections 9.1–9.5 502
Shortest-Processing-Time Scheduling 504
Earliest-Due-Date Scheduling 505
Minimizing the Number of Tardy Jobs 505
Precedence Constraints: Lawler’s A
­ lgorithm 506
nah77856_fm_new.indd 12
9.8
9.9
Snapshot Application: Millions Saved with
Scheduling System for Fractional Aircraft
Operators 508
Problems for Section 9.6 508
Sequencing Algorithms For Multiple
­Machines 510
Scheduling n Jobs on Two Machines 511
Extension to Three Machines 512
The Two-Job Flow Shop Problem 514
Problems for Section 9.7 517
Stochastic Scheduling: Static Analysis 518
Single Machine 518
Multiple Machines 519
The Two-Machine Flow Shop Case 520
Problems for Section 9.8 521
Stochastic Scheduling: Dynamic
­Analysis 522
Selection Disciplines Independent of Job
­Processing Times 524
Selection Disciplines Dependent on Job
­Processing Times 525
The cµ Rule 527
Problems for Section 9.9 527
9.10 Assembly Line Balancing 528
Problems for Section 9.10 532
Snapshot Application: Manufacturing
­Divisions Realize Savings with Scheduling
­Software 534
9.11 Historical Notes 535
9.12 Summary 535
Additional Problems on Scheduling 536
Bibliography 542
Chapter 10
Project Scheduling 543
Chapter Overview 543
10.1 Representing a Project as a Network 546
10.2 Critical Path Analysis 548
Finding the Critical Path 551
Problems for Sections 10.1 and 10.2 554
10.3 Time Costing Methods 556
Problems for Section 10.3 560
10.4 Solving Critical Path Problems with Linear
Programming 561
Linear Programming Formulation of the
Cost–Time Problem 564
Problems for Section 10.4 566
10.5 Pert: Project Evaluation and Review
­Technique 566
Path Independence 571
18/12/14 10:21 PM
Contents xiii
Problems for Section 10.5 574
Snapshot Application: Warner Robins
­Streamlines Aircraft Maintenance with
CCPM Project Management 576
10.6 Resource Considerations 576
Resource Constraints for Single-Project
­Scheduling 576
Resource Constraints for Multiproject
­Scheduling 578
Resource Loading Profiles 579
Problems for Section 10.6 581
11.7
10.10 Summary 587
Additional Problems on Project Scheduling 588
Appendix 10–A Glossary of Notation for
Chapter 10 591
11.8 Locating New Facilities 632
Snapshot Application: Kraft Foods Uses
­Optimization and Simulation to Determine
Best Layout 633
Measures of Distance 634
Problems for Section 11.8 635
11.9 The Single-Facility Rectilinear Distance
­Location Problem 635
Bibliography 592
Chapter 11
Facilities Layout and Location 593
Chapter Overview 593
11.1 The Facilities Layout Problem 596
11.2 Patterns of Flow 597
Activity Relationship Chart 598
From-To Chart 600
11.3 Types of Layouts 602
Fixed Position Layouts 602
Product Layouts 602
Process Layouts 603
Layouts Based on Group Technology 603
Problems for Sections 11.1–11.3 605
11.4
A Prototype Layout Problem and the ­
Assignment Model 607
The Assignment Algorithm 608
Problems for Section 11.4 610
*11.5 More Advanced Mathematical Programming
Formulations 611
11.10
11.11
Contour Lines 638
Minimax Problems 639
Problems for Section 11.9 642
Euclidean Distance Problems 643
The Gravity Problem 643
The Straight-Line Distance Problem 644
Problems for Section 11.10 645
Other Location Models 646
Locating Multiple Facilities 647
Further Extensions 648
Problems for Section 11.11 650
11.12 Historical Notes 651
11.13 Summary 652
Additional Problems on Layout and
­Location 653
Spreadsheet Problems for Chapter 11 658
Appendix 11–A Finding Centroids 659
Appendix 11–B Computing Contour Lines 661
Bibliography 644
Chapter 12
Quality and Assurance 666
Problem for Section 11.5 612
Chapter Overview 666
CRAFT 613
COFAD 617
12.1 Statistical Basis of Control Charts 671
11.6 Computerized Layout Techniques 612
nah77856_fm_new.indd 13
Flexible Manufacturing Systems 625
Advantages of Flexible Manufacturing
­Systems 627
Disadvantages of Flexible Manufacturing
Sysems 627
Decision Making and Modeling of the FMS 628
The Future of FMS 631
Problems for Section 11.7 632
10.7 Organizational Issues in Project
­Management 583
10.8 Historical Notes 584
10.9 Project Management Software for the PC 585
Snapshot Applications: Project Management
Helps United Stay on Schedule 586
Thomas Brothers Plans Staffing with Project
Management Software 586
Florida Power and Light Takes Project
­Management Seriously 586
ALDEP 618
CORELAP 619
PLANET 620
Computerized Methods versus Human
­ Planners 620
Dynamic Plant Layouts 621
Other Computer Methods 621
Problems for Section 11.6 622
Overview of this Chapter 670
Problems for Section 12.1 673
18/12/14 10:21 PM
xiv Contents
_
12.2 Control Charts for Variables: The X and R
Charts 675
_
X Charts 678
Relationship to Classical Statistics 678
R Charts 680
Problems for Section 12.2 681
Chapter 13
Reliability and Maintainability 742
Chapter Overview 742
13.1
Reliability of a Single Component 746
12.3 Control Charts for Attributes: The p Chart 683
p Charts for Varying Subgroup Sizes 685
Problems for Section 12.3 686
Introduction to Reliability Concepts 746
Preliminary Notation and Definitions 747
The Exponential Failure Law 749
Problems for Section 13.1 752
Increasing and Decreasing Failure Rates 754
12.4 The c Chart 688
13.2
12.5 Classical Statistical Methods and Control
Charts 691
Problem for Section 12.5 691
_
*12.6 Economic Design of X Charts 692
13.3
The Poisson Process in Reliability
­Modeling 757
12.7 Overview of Acceptance Sampling 699
13.4 Failures of Complex Equipment 762
Problems for Section 12.4 690
Problems for Section 12.6 698
Problems for Section 13.2 756
Series Systems Subject to Purely Random
­ Failures 760
Problems for Section 13.3 761
Components in Series 762
Components in Parallel 763
Expected Value Calculations 763
K out of N Systems 764
Problems for Section 13.4 766
Snapshot Application: Navistar Scores with
­Six-Sigma Quality Program 701
12.8 Notation 702
12.9 Single Sampling for Attributes 702
Derivation of the OC Curve 704
Problems for Section 12.9 706
*12.10 Double Sampling Plans for Attributes 707
Problems for Section 12.10 708
12.11 Sequential Sampling Plans 709
12.12
12.13
13.5 Introduction to Maintenance Models 766
13.6 Deterministic Age Replacement
­Strategies 768
The Optimal Policy in the Basic Case 768
A General Age Replacement Model 770
Problems for Section 13.6 774
Problems for Section 12.11 713
Average Outgoing Quality 714
Snapshot Application: Motorola Leads the
way with Six-Sigma Quality Programs 716
Problems for Section 12.12 716
13.7 Planned Replacement under Uncertainty 774
Definitions 717
Listening to the Customer 717
Competition Based on Quality 719
Organizing for Quality 720
Benchmarking Quality 721
The Deming Prize and the Baldrige Award 722
ISO 9000 724
Quality: The Bottom Line 725
*13.8
Design, Manufacturing, and Quality 728
13.10 Historical Notes 789
13.11 Summary 790
Total Quality Management 717
12.14 Designing Quality into the Product 726
12.15 Historical Notes 730
12.16 Summary 731
Additional Problems on Quality and
­Assurance 733
Appendix 12–A Approximating
Distributions 737
Appendix 12–B Glossary of Notation for
Chapter 12 on Quality and Assurance 739
Bibliography 740
nah77856_fm_new.indd 14
Planned Replacement for a Single Item 774
Block Replacement for a Group of Items 778
Problems for Section 13.7 780
Analysis of Warranty Policies 782
The Free Replacement Warranty 782
The Pro Rata Warranty 784
Extensions and Criticisms 786
Problems for Section 13.8 786
13.9 Software Reliability 787
Snapshot Application: Reliability-Centered
­Maintenance Improves Operations at Three
Mile Island Nuclear Plant 788
Additional Problems on Reliability and
­Maintainability 791
Appendix 13–A Glossary of Notation on
Reliability and Maintainability 793
Bibliography 795
Appendix: Tables 796
Index 814
18/12/14 10:21 PM
About the Authors
Steven Nahmias is Professor of Operations Management and Information Systems
in the Leavey School of Business Administration at Santa Clara University. He holds
a B.A. in Mathematics and Sciences from Queens College, and B.S. in Operations
Research from Columbia University, and the M.S. and Ph.D. degrees in Operations
Research from Northwestern. He previously served on the faculties of the University
of Pittsburgh, Georgia Tech, and Stanford University.
He is widely known for his research on stochastic inventory models with an ­emphasis
on perishables. He has authored or co-authored more than 50 scientific articles which
have appeared in a variety of national and international journals. He has served as Area
Editor in Supply Chain Logistics for Operations Research, senior editor for M&SOM,
associate editor for Management Science, and associate editor for Naval Research
Logistics. He earned first prize in the Nicholson Student Paper competition (1971) and
second prize in the TIMS student paper competition (1972) and the University Award
for Sustained Excellence in Scholarship from Santa Clara University (1998). In 2011
he was named Distinguished Fellow of the Manufacturing and Service Operations
Management Society and in 2014 Distinguished Fellow of INFORMS.
In addition to his academic activities he has served as a consultant to a variety
of companies and agencies. In his spare time he enjoys biking and golf. He is also a
semi-professional jazz trumpet player and performs regularly with several bands in
the Bay Area.
Tava Olsen holds the Ports of Auckland chair in Logistics and Supply Chain
­Management at the University of Auckland Business School. Prior to joining ­Auckland,
she was Professor of Operations and Manufacturing Management in the Olin Business
School at Washington University in St. Louis, which she joined after serving as an
Assistant Professor in the Department of Industrial and Operations Engineering at the
University of Michigan, Ann Arbor. Dr. Olsen received her B.Sc. (honours) in Mathematics from the University of Auckland. She earned both her M.S. in Statistics and
her Ph.D. in Operations Research from Stanford University.
Dr. Olsen’s research interests include supply chain management; pricing and
­inventory control; and stochastic modeling of manufacturing, service, and healthcare systems. Among other journals, her publications have appeared in Management
­Science, Operations Research, Manufacturing and Service Operations Management
(M&SOM), and the Journal of Applied Probability. In 2012, she received Auckland
Business School’s Sustained Research Excellence Award.
Dr. Olsen is the Academic Director of the Centre for Supply Chain ­Management.
She is currently an Associate Editor for Management Science, M&SOM and O
­ perations
Research, is a senior editor of Production and Operations Management. She is a past
president of the Manufacturing and Service Operations (MSOM) society for which
she received the society’s distinguished service award. She enjoys hiking, biking, and
spending time with her husband and two daughters.
xv
nah77856_fm_new.indd 15
18/12/14 10:21 PM
Preface to the Seventh Edition
This edition is the most substantial to date. All of the changes have been to the first
half of the book, which appears to comprise the most popular chapters. Chapter 1 is
basically intact with the addition of a major section on servicization. In Chapter 2,
the tables that appeared in the first six editions have been replaced with screen shots
of spreadsheets with cell formula definitions. The section on seasonal decomposition
has been streamlined, and several new snapshot applications were added. Chapter 3
has been retitled and repositioned. The chapter is now titled “Sales and Operations
­Planning.” It incorporates the aggregate planning material from prior editions, but
­includes new material on the planning process and discusses how production plans
grow out of demand forecasts. Chapter 4 on deterministic demand includes a new
Snapshot Application and a brief case study at the end of the chapter. In Chapter 5,
there is a new appendix that derives a closed form expression for the expected cost
function in a newsvendor model with normal demand. Chapter 6 on supply chains has
been substantially revised and repositioned. Chapter 7 is new. It presents a comprehensive treatment of the analytical models for service systems and incorporates material
on queueing and the Poisson process from the prior editions. The remaining chapters
(8 through 13) are largely the same as the sixth edition.
Finally, I am very happy to welcome Tava Lennon Olsen as a co-author. She has
brought a fresh perspective to the book and broadened the coverage substantially.
Steven Nahmias
xvi
nah77856_fm_new.indd 16
18/12/14 10:21 PM
Chapter One
Strategy and
­Competition
“However beautiful the strategy, you should occasionally look at the results.”
—Winston Churchill
Chapter Overview
Purpose
The purpose of this chapter is to introduce the student to a variety of strategic
­issues that arise in the manufacturing function of the firm.
Key Points
1. Manufacturing matters. This writer contends that the loss of the manufacturing
base in the U.S. economy is not healthy and will eventually lead to an overall loss
in the standard of living and quality of life in this country. It counters the argument
that our evolution into a service economy is a natural and healthy thing.
2. Strategic dimensions. Along with cost and/or product differentiation, other
­dimensions along which firms distinguish themselves include (a) quality,
(b) delivery speed, (c) delivery reliability, and (d) flexibility.
3. Global competition. How do we measure our success and economic health on
a global scale? One way is to examine classical measures of relative economic
strength, which include (a) balance of trade, (b) share of world exports,
(c) creation of jobs, and (d) cost of labor. However, such macro measures do not
adequately explain why certain countries dominate certain industries. National
competitive advantage is a consequence of several factors (factor conditions,
­demand conditions, related and supporting industries, firm strategy structure,
and rivalry), although productivity also plays an important role.
4. Strategic initiatives. We discuss several strategic initiatives that have allowed
many companies to shine in their respective arenas. These include (a) business
process reengineering, (b) just-in-time manufacturing and purchasing systems,
(c) time-based competition, and (d) competing on quality.
5. Product and process life cycles. Most of us understand that products have
natural life cycles: start-up, rapid growth, maturation, stabilization, or decline.
However, it is rarely recognized that processes too have life cycles. Initially, new
manufacturing processes have the characteristics of a job shop. As the process
matures, automation is introduced. In the mature phases of a manufacturing
process, most major operations are automated. A firm needs to match the phases
of product and process life cycles to be the most successful in its arena.
1
nah77856_ch01.indd 1
17/12/14 6:02 PM
2 Chapter One Strategy and Competition
6. Learning and experience curves. These are helpful in forecasting the decline in
unit cost of a manufacturing process as one gains experience with the process.
Learning curves are more appropriate when modeling the learning of an
individual worker, and experience curves are more appropriate when considering
an entire industry.
7. Capacity growth planning. Another important strategic issue in operations is
­determining the timing and sizing of new capacity additions. Simple models
(make or buy problem) and more complex exponential growth models are
explored in Section 1.11. In addition, some of the factors that determine
­appropriate location of new facilities is explored.
Strategy is a long-term plan of action designed to achieve a particular goal, most
often winning. Its root is from the Greek strate–gos, which referred to a “military commander” during the age of Athenian Democracy. Strategy was originally conceived in
the military context. Two famous books dealing with military strategy are The Prince
by Machiavelli and The Art of War by Sun Tzu.
Hence, we can see that business strategy relates closely to military strategy. Companies fight on an economic battlefield, and long-term strategies determine winners
and losers. Business strategy is the highest level of corporate activity that bundles
together the disparate functional area strategies. Business strategy sets the terms and
goals for a company to follow.
Perhaps the reason that chief executive officers (CEOs) are compensated so highly
in the United States is the realization that the strategic vision of the CEO is often the
difference between the success and failure of a company. The strategic visions of industry giants such as Henry Ford, Jack Welch, and Bill Gates were central to the success of their companies that have, at one time or another, dominated their competition.
Perhaps the most dramatic example is Apple Corporation. With the introduction of
the iPod in 2002 and the iPhone is 2007, Apple transformed itself from a failing computer company to a major force in portable computing and telecommunications. The
fascinating transformation of the firm is described in the Snapshot Application on the
next page.
Success requires a vision, and visions must be articulated so all of the firm’s
­employees can share in that vision. The formal articulation of the vision is known
as the company mission statement. A good mission statement should provide a clear
description of the goals of the firm and is the first step toward formulating a coherent business strategy. Poor mission statements tend to be wordy and full of generalities. Good mission statements are direct, clear, and concise. In their book, Jones and
­Kahaner (1995) list the 50 corporate mission statements that they perceive as the best.
One example is the Gillette Corporation. Their mission statement is: “Our Mission is to
achieve or ­enhance clear leadership, worldwide, in the existing or new core consumer
product categories in which we choose to compete.” They then go on to list ­exactly
which areas they perceive as their core competencies. Intel defines their ­mission as:
“Do a great job for our c­ ustomers, employees, and stockholders by being the preeminent building block s­ upplier to the computing industry.” The statement then provides
­details on “Values and Objectives.” In many cases, their objectives are quite specific
(e.g., “Lead in LAN products and Smart Network Services”). Certainly, the award for
conciseness has to go to the General Electric Corporation, whose mission statement
is three words: “Boundaryless . . . Speed . . . Stretch.” The commentary ­following the
statement provides an explanation of exactly what these words mean in the context of
the corporation.
nah77856_ch01.indd 2
17/12/14 6:02 PM
Snapshot Application
Apple Adopts a New Business Strategy
and Shifts Its Core Competency from
Computers to Portable Electronics
Apple Computer was the kind of success story one sees
in the movies. Two youngsters, Steve Wozniak and Steve
Jobs, grew up with an interest in hobbyist computers.
Working from a garage, they founded Apple Computer
in April 1976, and soon after, introduced a build your
own hobbyist computer called the Apple I. The firm was
incorporated a year later with the help of Mike Markula,
and the Apple II was introduced in April 1977, ushering in the world of personal computing. Perhaps it was
the fact that it was selected as the platform for Visicalc,
the first spreadsheet program, that led to its success, as
much as the superior capabilities of the hardware.
While personal computers have become a common part
of our everyday lives, we forget that they are a relatively
new invention. The nature of the personal computer marketplace was dramatically altered by the introduction of
the first PC by IBM in 1981. IBM’s open architecture allowed
for inexpensive clones to enter the market, and crowd out
Apple’s significantly more expensive products. By the turn
of the century, Apple’s future looked to be in doubt.
Apple’s subsequent transformation and rebirth is
a fascinating bit of business history. Around 2001, an
independent consultant, Tony Fadell, was shopping
around his concept of an MP3 music player linked to a
music sales service. At that time, MP3 players were not
new; one could “rip” music from one’s CD collection,
and load the songs on the player. While no one else was
interested in Fadell’s idea, he was hired by Apple and
assigned to a team of 30 people, including designers,
programmers, and hardware engineers.
When Apple decided to go ahead with the MP3
player concept, they also decided that they needed a
new design to separate themselves from the rest of the
marketplace. Apple subcontracted much of the development and design work to PortalPlayer, who devoted
all of their resources to the project. Steve Jobs himself
was intimately involved with the design and function
of the new player. The iPod was a huge success and has
been redesigned over several models and generations.
The most current reports record worldwide sales of over
350 million units.
While the iPod was a huge success, Apple did not rest
on its laurels. In 2007, Apple launched the first iPhone.
Again, the concept of a smartphone was not new. Several companies, notably Motorola, Samsung, Palm, and
Nokia, had smartphones on the market for several years
prior. But as with the iPod, Apple again produced an
innovative product with unique features. Apple continues to improve upon the iPhone and introduces a new
generation of the product virtually every year. As of this
writing, sales have reached over 500 million worldwide.
Apple’s most recent product, the iPad, was also an
instant success and essentially defined a new market category. This tablet computer, introduced in March 2010,
is convenient for web surfing and reading e-books, and
again, has become the product the competition measures itself against. Apple registered more than 1 million
sales of the iPad in the first three months alone. As a
testament to Apple’s phenomenal success in portable
computing, the market capitalization of Apple surpassed
that of the software behemoth Microsoft in 2010.
Source: Various websites and L. Kahney “Inside Look at the
Birth of the iPod” July 2004 (http:/www.wired.com).
Once having articulated a vision, the next step is to plan a strategy for achieving that
vision. This is the firm’s business strategy. The overall business strategy includes defining
1.
2.
3.
4.
The market in which the enterprise competes.
The level of investment.
The means of allocating resources to and the integrating of separate business units.
Functional area strategies, including
• The marketing strategy
• The financial strategy
• The operations strategy
Broadly defined, the operations strategy is the means by which the firm deploys
its ­resources to achieve its competitive goals. For manufacturing firms, it is the sum
total of all decisions concerning the production, storage, and distribution of goods.
­Important operations strategy decisions include where to locate new manufacturing
­facilities, how large these facilities should be, what processes to use for manufacturing
nah77856_ch01.indd 3
17/12/14 6:02 PM
nah77856_ch01.qxd
4
12/28/07
3:46 PM
Page 4
nah77856
Chapter One Strategy and Competition
and moving goods through the system, and what workers to employ. Service firms also
require an operations strategy. The United States continues to be a leader in financial
services, which must be supported by effective and reliable operations departments in
order to remain competitive. The Disney theme park’s continuing record of success is
due in part to its careful attention to detail in every phase of its operations.
Does the American culture place too much emphasis on marketing (selling the product) and finance (leveraged buyouts, mergers, stock prices) and too little on operations
(making and delivering the product)? Years ago, this was certainly the case. However,
we are quick learners. The enormous success of the Japanese auto industry, for example, provided strong motivation for the American big three to close their inefficient
plants and change the way things were done. The dramatic differences that were
brought to light by Womack, Jones, and Roos (1990) have largely been eliminated.
Today, the best American auto plants rival their Japanese counterparts for quality and
efficiency.
Still, a coherent operations strategy is essential. When the Apple Macintosh was introduced, the product was extremely successful. However, the company was plagued
with backorders and failed to keep up with consumer demand. According to Debbi
Coleman, Apple’s former director of worldwide manufacturing:
Manufacturing lacked an overall strategy which created problems that took nine months to
solve . . . we had extremely poor forecasting. Incoming materials weren’t inspected for defects
and we didn’t have a mechanism for telling suppliers what was wrong, except angry phone
calls. Forty percent of Mac materials were coming from overseas and no one from Apple
was inspecting them before they were shipped. . . . One of the biggest tasks that high-tech
manufacturers face is designing a manufacturing strategy that allows a company to be flexible
so it can ride with the highs and lows of consumer and business buying cycles. (Fallon, 1985)
Although it is easy to be critical of American management style, we must be aware
of the factors motivating American managers and those motivating managers from
other cultures. For example, the Japanese have not achieved their dramatic successes
without cost. Sixteen-hour work days and a high rate of nervous breakdowns among
management are common in Japan.
Measuring a firm’s success by the performance of its share price can result in shortsighted management practices. Boards of directors are more concerned with the next
quarterly report than with funding major long-term projects. In fact, Hayes and
Wheelwright (1984) make a compelling argument that such factors led to a myopic
management style in the United States, characterized by the following:
1. Managers’ performance is measured on the basis of return on investment (ROI),
which is simply the ratio of the profit realized by a particular operation or project
over the investment made in that operation or project.
2. Performance is measured over short time horizons. There is little motivation for a
manager to invest in a project that is not likely to bear fruit until after he or she has
moved on to another position.
In order to improve ROI, a manager must either increase the numerator (profits) or decrease the denominator (investment). In the short term, decreasing the denominator by
cutting back on the investment in new technologies or new facilities is easier than trying
to increase profits by improving efficiency, the quality of the product, or the productivity
of the operating unit. The long-term effects of decreasing investment are devastating. At
some point, the capital costs required to modernize old factories become more than the
firm can bear, and the firm loses its competitive position in the marketplace.
nah77856_ch01.indd 4
17/12/14 6:02 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 5
1.1
Manufacturing Matters 5
It would be encouraging if the problems of U.S. industries arising from overemphasis on short-term financial performance were decreasing, but sadly, they appear to be
worsening. Because of gross mismanagement and questionable auditing practices, two
giants of American industry were brought down in 2001: Enron and Arthur Andersen.
“Enron went from the No. 7 company on the Fortune 500 to a penny stock in a stunning three weeks because it apparently lied on its financial statements,” said
Representative John D. Dingell, one-time member of the House Energy Committee.
While other parts of the world have experienced spectacular problems as well (such as
the Asian financial crisis that hit in the late 1990s), few Americans can understand how
a company that had recently expanded and profited from the energy crisis, and an
American icon such as Arthur Andersen, could both be brought down so quickly and
completely. It is our continual focus on short-term performance and the incentive
system we have built up around this objective that led to these crises.
Measuring individual performance over the short term is a philosophy that seems to
pervade American life. Politicians are elected for two-, four-, or six-year terms. There
is a strong incentive for them to show results in time for the next election. Even university professors are evaluated yearly on their professional performance in many
institutions, even though most serious academic projects extend over many years.
1.1
MANUFACTURING MATTERS
A question that is being debated and has been debated by economists for several
decades is the importance of a strong manufacturing base. The decline of manufacturing domestically has led to a shift in jobs from the manufacturing sector to the service
sector. Because there are major disparities in labor costs in different parts of the world,
there are strong incentives for American firms to locate volume manufacturing facilities overseas to reduce labor costs. Is a strong manufacturing base important for the
health of the economy?
There is little debate that manufacturing jobs have been steadily declining in the
United States. The growth of manufacturing overseas, and in China in particular, is
well documented. If we compare the proportion of nonagriculture jobs in the United
States in service versus manufacturing in 1950 versus 2002, the change is quite
dramatic. In 1950, manufacturing jobs accounted for 34 percent of nonagriculture
labor and service jobs accounted for 59 percent. In 2002, however, manufacturing jobs
only accounted for 13 percent of nonagriculture jobs, while service jobs soared to
82 percent of the total (Hagenbaugh, 2002).
One mitigating factor in the loss of manufacturing was the dramatic rise in manufacturing productivity during this same period. Average annual manufacturing productivity
growth was 2.57 percent annually during the 1980s and 3.51 percent annually during the
1990s (Faux, 2003). This dramatic rise in manufacturing productivity has had the effect
of offsetting the loss of high-paying manufacturing jobs at home, thus partially accounting for the success of the U.S. economy in the latter part of the first decade of the century.
An argument put forth by several scholars (e.g., Daniel Bell, 1976) is that we are
simply evolving from an industrial to a service economy. In this view, the three stages
of economic evolution are (1) agrarian, (2) industrial, and (3) service. In the early years
of our country, we were primarily an agrarian economy. With the industrial revolution,
a large portion of the labor force shifted from agriculture to manufacturing. In recent
years it seems that there is less interest in manufacturing. These scholars would argue
nah77856_ch01.indd 5
17/12/14 6:02 PM
6 Chapter One Strategy and Competition
that we are merely entering the third stage of the evolutionary process: moving from
an industrial economy to a service economy.
It is comforting to think that the American economy is healthy and simply evolving from an industrial to a service economy. One might even argue that manufacturing is
not important for economic well-being. According to economist Gary S. Becker (1986),
“Strong modern economies do not seem to require a dominant manufacturing sector.’’
It is far from clear, however, that we evolved from an agrarian economy to an
­industrial economy. Although fewer American workers are employed in the agricultural sector of the economy, agricultural production has not declined. Based on U.S.
Department of Commerce data, Cohen and Zysman (1987) state that “agriculture has
sustained, over the long term, the highest rate of productivity increase of any sector.’’
By utilizing new technologies, agriculture has been able to sustain growth while consuming fewer labor hours. Hence, the figures simply do not bear out the argument that
our economy has shifted from an agricultural one to an industrial one.
The argument that the economy is undergoing natural stages of evolution is simply
not borne out by the facts. I believe that all sectors of the economy—agricultural,
manufacturing, and service—are important and that domestic economic well-being
­depends upon properly linking the activities of these sectors.
The return on innovations will be lost if new products are abandoned after development. The payoff for research and development (R&D) can come only when the product
is produced and sold. If manufacturing is taken offshore, then the “rent on innovation’’
cannot be recaptured. Furthermore, manufacturing naturally leads to innovation. It will
be difficult for the United States to retain its position as a leader in innovation if it loses
its position as a leader in manufacturing.
That manufacturing naturally leads to innovation is perhaps best illustrated by the
Japanese experience in the video market. After Japan had captured the lion’s share of
the world market for televisions, the next major innovation in consumer video technology, the videocassette recorder (VCR), (at least, the inexpensive consumer version)
was developed in Japan, not the United States. Virtually all VCRs sold were manufactured in Asia.
A more recent book by Pisano and Shih (2012) underscores many of the same
themes that appeared in Cohen and Zysman (1987). They point to other products that
were invented in the United States but whose base of manufacturing is now overseas.
An example is the PV (photovoltaic) cell (more commonly known as the solar cell).
PV cells were invented in Bell Labs, but only a very small percentage of the world
demand is filled by American companies.
One of the more disturbing trends discussed by Pisano and Shih (2012) is the widening trade deficit in manufactured goods. The foreign trade deficit has continued to
increase, resulting in the United States going from the largest creditor nation in the
1970s to the largest debtor nation today. The overwhelming source of this deficit is the
continued negative trade balance in manufactured goods (the trade balance in services
is actually increasing).
Where to locate manufacturing as well as R and D facilities is one of the key
management decisions a firm must make. During the 1990s we saw an exodus of
domestic manufacturing to China. The offshoring movement was rampant, not only
in the United States but in most developed countries. The primary driver of this exodus is wage rates, but other factors were relevant as well. Favorable tax treatments,
proximity to natural resources, and proximity to markets are also reasons companies
locate facilities offshore.
In recent years, the advantage of offshoring is decreasing. For example, as the
standard of living in China has improved, manufacturing wage rates have risen. When
nah77856_ch01.indd 6
17/12/14 6:02 PM
1.1 Manufacturing Matters 7
the disadvantages of offshoring are taken into account, the best course of action is no
longer obvious. These disadvantages include longer lead times, infrastructure deficiencies, local politics, and quality problems.
An example cited in Reshoring Manufacturing (2013) is the start-up company
ET Water Systems. In 2005 the firm moved manufacturing operations to China in
search of lower labor costs. However, the disadvantages of locating offshore became
apparent as the company starting suffering losses due to several factors, including
the cost of funds tied up in goods in transit, the disconnect between manufacturing
and design, and recurring quality problems. When the firm’s chief executive, Mark
Coopersmith, carefully looked at the total cost difference between manufacturing in
China versus California he was amazed to discover that California was only 10%
more expensive than China. He concluded that this cost difference was more than
offset by the advantages of locating manufacturing domestically. ET Water Systems
closed their plant in China and reshored the manufacturing function to General Electronics Assembly in San Jose.
Unfortunately, ET’s experience is rare. More companies are continuing to choose
the offshoring option. However, that reshoring is occurring at all is a positive step.
Perhaps more companies will come to the same conclusion that ET Water Systems did
when taking into account the full spectrum of the costs of offshoring.
In order to get some idea of the extent of reshoring compared to offshoring, Porter
and Rivkin (2012) conducted an extensive survey of Harvard Business School alumni
who made location decisions for their companies in 2011. They found that only 9% of
the respondents were considering moving offshore activities back to the United States,
34% were planning on keeping their facilities where they were, and 57% were considering moving existing facilities out of the United States. Their results suggest that
offshoring still dominates both staying put and reshoring. The respondents’ main reasons for offshoring were lower wages, proximity to customers, better access to skilled
labor, higher labor productivity, lower taxes, proximity of suppliers, and proximity to
other company operations. The respondents’ main reasons for staying put or reshoring
were proximity to the U.S. market, less corruption, greater safety, better intellectual
property protection, similar language and culture, better infrastructure, and proximity
to other company operations.
Has offshoring been a help or a hindrance to the U.S. economy? The answer is not
­simple. On one hand, offshoring has resulted in loss of jobs domestically, lower average
­domestic wages which in turn have yielded a lower tax base and a smaller domestic market. On the other hand, offshoring has improved the bottom line for many domestic firms,
and have resulted in lower costs of manufactured goods for the American consumer.
Manufacturing Jobs Outlook
The U.S. Bureau of Labor Statistics (a subsidiary of the Department of Labor) p­ rovides
up-to-date information on the prospects for jobs in the manufacturing sector by
industry. According to the Occupational Outlook Handbook (OOH), 2010–2011 Edition
(http://www.bls.gov/oco/), even though manufacturing jobs are expected to decline
overall, there are some areas of growth and opportunity. Consider the individual ­sectors:
1. Aerospace products and parts. This sector is projected to grow, but more slowly
than the economy in general. Earnings are higher here than in most other manufacturing industries, as workers must be highly skilled. Opportunities will result from
a large number of anticipated retirements.
2. Chemical (except pharmaceuticals and medicines). The chemical industry continues to be a major employer of professionals, producing over 500,000 jobs.
nah77856_ch01.indd 7
17/12/14 6:02 PM
8 Chapter One Strategy and Competition
3.
4.
5.
6.
7.
8.
9.
10.
However, employment is projected to decline and competition for better jobs to
increase over the coming years.
Computer and Electronic Products. Employment is projected to decrease nearly
20 percent in the decade 2008–2018 due to productivity improvements and movement of jobs to lower wage countries.
Food Manufacturing. The jobs picture in this industry is stable, but production workers continue to have the highest incidences of injury and illness among all industry,
with seafood product preparation and packaging being the worst sector in this regard.
Machinery. Productivity improvements will lead to fewer jobs overall, but opportunities will arise as a result of anticipated retirements. Machinery manufacturing has
some of the most highly skilled, and highly paid, production jobs in manufacturing.
Motor Vehicles and Parts. Almost half the jobs are located in Michigan, Indiana,
and Ohio, but jobs continue to shift away from this area to the South. Average
earnings continue to be high in this sector, but employment is expected to decline
in coming years.
Pharmaceuticals and Medicine. This continues to be a growth area, with earnings higher than in other manufacturing industries. Job prospects are particularly
favorable for candidates with advanced degrees.
Printing. Most printing establishments are very small, with 70 percent employing
under 10 people. Traditional printing is a declining industry due to increased computerization, but digital press operators will continue to be in demand.
Steel. Steel continues to be a declining industry domestically, with fewer jobs projected as a result of consolidation and automation. Opportunities will be best for
engineers and skilled production and maintenance workers.
Textiles, Textile Products, and Apparel. About half the jobs are located in three
states: California, North Carolina, and Georgia. Employment is expected to d­ ecline
rapidly because of technological advances and imports of apparel and textiles from
lower wage countries.
1.2 A Framework for Operations Strategy
Classical literature on competitiveness claims that firms position themselves strategically in the marketplace along one of two dimensions: lower cost or product differentiation (Porter, 1990).
Often new entrants to a market position themselves as the low-cost providers. Firms
that have adopted this approach include the Korean automakers (Hyundai, Daewoo, Kia),
discount outlets such as Costco, and retailers such as Wal-Mart. While being the low-cost
provider can be successful over the near term, it is a risky strategy. Consumers ultimately
will abandon products that they perceive as poor quality regardless of cost. For example,
many manufacturers of low-cost PC clones popular in the 1980s are long gone.
Most firms that have a long record of success in the marketplace have differentiated
themselves from their competitors. By providing uniqueness to buyers, they are able
to sustain high profit margins over time. One example is BMW, one of the most profitable auto firms in the world. BMW continues to produce high-performance, well-made
cars that are often substantially more expensive than those of competitors in their class.
­Product differentiation within a firm has also been a successful strategy. Consider the
­success of General Motors in the early years compared to Ford. GM was able to successfully capture different market segments at the same time by forming five distinct
nah77856_ch01.indd 8
17/12/14 6:02 PM
1.2 A Framework for Operations Strategy 9
divisions, while Henry Ford’s insistence on providing only a single model almost led
the company to bankruptcy (Womack et al., 1990).
Strategic Dimensions
However, cost and product differentiation are not the only two dimensions along which
firms distinguish themselves. The following additional factors relate directly to the
­operations function:
•
•
•
•
Quality
Delivery speed
Delivery reliability
Flexibility
What does quality mean? It is a word often bandied about, but one that means different things in different contexts. Consider the following hypothetical remarks.
1. “That hairdryer was a real disappointment. It really didn’t dry my hair as well as I
expected.”
2. “I was thrilled with my last car. I sold it with 150,000 miles and hardly had any r­ epairs.”
3. “I love buying from that catalogue. I always get what I order within two days.”
4. “The refrigerator works fine, but I think the shelves could have been laid out b­ etter.”
5. “That park had great rides, but the lines were a mess.”
6. “Our quality is great. We’ve got less than six defectives per one million parts p­ roduced.”
In each case, the speaker is referring to a different aspect of quality. In the first case,
the product simply didn’t perform the task it was designed to do. That is, its function was
substandard. The repair record of an automobile is really an issue of reliability rather
than quality, per se. In the third case, it is delivery speed that translates to quality service
for that customer. The fourth case refers to a product that does what it is supposed to do,
but the consumer is disappointed with the product design. The product quality (the rides)
at the amusement park were fine, but the logistics of the park management were a disappointment. The final case refers to the statistical aspects of quality control.
Hence the word quality means different things in different contexts. A Honda Civic
is a quality product and so is a Ferrari Testarosa. Consumers buying these products are
both looking for quality cars but have fundamentally different objectives. The fact is
that everyone competes on quality. For this reason, Terry Hill (1993) would classify
quality as an order qualifier rather than an order winner. An option is immediately
eliminated from consideration if it does not meet minimum quality standards. It is the
particular aspect of quality on which one chooses to focus that determines the ­nature
of the competitive strategy and the positioning of the firm.
Delivery speed can be an important competitive weapon in some contexts. Some
firms base their primary competitive position on delivery speed, such as UPS and
­Federal Express. Mail-order and Web-based retailers also must be able to deliver products reliably and quickly to remain competitive. Building contractors that complete
projects on time will have an edge.
Delivery reliability means being able to deliver products or services when pro­m­
ised. Online brokerages that execute trades reliably and quickly will retain customers.
Contract manufacturers are measured on several dimensions, one being whether they
can deliver on time. As third-party sourcing of manufacturing continues to grow, the
successful contract manufacturers will be the ones that put customers first and maintain a record of delivering high-quality products in a reliable fashion.
nah77856_ch01.indd 9
17/12/14 6:02 PM
10 Chapter One Strategy and Competition
Flexibility means offering a wide range of products and being able to adjust to
u­ nexpected changes in the demand of the product mix offered. Successful manufacturers
in the 21st century will be those that can respond the fastest to unpredictable changes in
customer tastes. This writer was fortunate enough to tour Toyota’s M
­ otomachi Plant located in Toyoda City, Japan. What was particularly impressive was the ability to produce
several different models in the same plant. In fact, each successive car on the a­ ssembly
line was a different model. A right-hand drive Crown sedan, for the d­ omestic market,
was followed by a left-hand drive Lexus coupe, designated for shipment to the United
States. Each car carried unique sets of instructions that could be read by both robot welders and human assemblers. This flexibility allowed Toyota to adjust the ­product mix on
a real-time basis and to embark on a system in which customers could order customconfigured cars directly from terminals located in dealer showrooms (Port, 1999).
Hence, one way to think of operations strategy is the strategic positioning the firm
chooses along one of the dimensions of cost, quality, delivery speed, delivery reliability,
and flexibility. Operations management is concerned with implementing the strategy
to achieve leadership along one of these dimensions.
1.3 Competing in the Global Marketplace
International competitiveness has become a national obsession. Americans are con­
cerned that their standard of living is eroding while it seems to improve elsewhere.
Evidence exists that there is some truth to this perception. Our balance of trade with
Japan has been in the red for decades, with no evidence of a reversal. American firms
once held a dominant position worldwide in industries that have nearly disappeared
domestically. Consumer electronics, steel, and machine tools are some examples. All
the news is not bad, however. The American economy is strong and continues to grow.
American firms still have the lion’s share of the world market in many industries.
In his excellent study of international competitiveness, Porter (1990) poses the following question: Why does one country become the home base for successful international competitors in an industry? That certain industries flourish in certain countries
cannot be disputed. Some examples are
1.
2.
3.
4.
5.
Germany: printing presses, luxury cars, chemicals.
Switzerland: pharmaceuticals, chocolate.
Sweden: heavy trucks, mining equipment.
United States: personal computers, software, films.
Japan: automobiles, consumer electronics, robotics.
What accounts for this phenomenon? One can offer several compelling explanations,
but most have counterexamples. Here are a few:
1. Historical. Some industries are historically strong in some countries and are not
easily displaced. Counterexample: The demise of the steel industry in the United
States is one of many counterexamples.
2. Tax structure. Some countries, such as Germany, have no capital gains tax, thus
providing a more fertile environment for industry. Counterexample: However, there
is no reason that favorable tax treatment should favor certain industries over others.
3. National character. Many believe that workers from other countries, particularly from
Pacific Rim countries, are better trained and more dedicated than American workers.
Counterexample: If this is true, why then do American firms dominate in some industry
nah77856_ch01.indd 10
17/12/14 6:02 PM
1.3 Competing in the Global Marketplace 11
4.
5.
6.
7.
8.
segments? How does one explain the enormous success Japanese-based corporations
have had running plants in the United States with an American workforce?
Natural resources. There is no question that some industries are highly ­resource
­dependent and these industries have a distinct advantage in some countries. One
­example is the forest products industry in the United States and Canada. Counterexample: Many industry sectors are essentially resource independent but still seem to
flourish in certain countries.
Government policies. Some governments provide direct assistance to fledgling industries, such as MITI in Japan. The role of the U.S. government is primarily regulatory.
For example, environmental standards in the United States are probably more stringent than almost anywhere else. Counterexample: This does not explain why some
­industries dominate in countries with strict environmental and regulatory standards.
Advantageous macroeconomic factors. Exchange rates, interest rates, and government
debt are some of the macroeconomic factors that provide ­nations with competitive advantage. For example, in the 1980s when interest rates were much higher in the United
States than they were in Japan, it was much easier for Japanese firms to borrow for new
projects. Counterexample: These factors do not explain why many nations have a r­ ising
­standard of living despite rising deficits (Japan, Italy, and Korea are some examples).
Cheap, abundant labor. Although cheap labor can attract new industry, most countries with cheap labor are very poor. On the other hand, many countries (Germany,
Switzerland, and Sweden are examples) have a high standard of living, high wage
rates, and shortages of qualified labor.
Management practices. There is evidence that Japanese management practices are
more effective in general than Western-style practices. Counter­example: If American
management practices are so ineffective, why do we continue to dominate certain
­industries, such as personal computers, software development, and pharmaceuticals?
Talking about competitiveness is easier than measuring it. What are the appropriate
ways to measure one country’s success over another? Some possibilities are
•
•
•
•
Balance of trade.
Share of world exports.
Creation of jobs.
Low labor costs.
Arguments can be made against every one of these as a measure of international competitiveness. Switzerland and Italy have trade deficits, and at the same time have experienced rising standards of living. Similar arguments can be made for countries that import more than they export. The number of jobs created by an economy is a poor gauge
of the health of that economy. More important is the quality of the jobs created. Finally,
low labor costs correlate with a low standard of living. These counterexamples show
that it is no easy task to develop an effective measure of international competitiveness.
Porter (1990) argues that the appropriate measure to compare national performance
is the rate of productivity growth. Productivity is the value of output per unit of input
of labor or capital. Porter argues that productivity growth in some industries ­appears
to be stronger in certain countries, and that there are reasons for this. In some cases we
can find the reasons in domestic factor advantages. The factor theory says all countries
have access to the same technology (an assumption that is not strictly true) and that
­national advantages accrue from endowments of production factors such as land, labor,
natural resources, and capital.
nah77856_ch01.indd 11
17/12/14 6:02 PM
12 Chapter One Strategy and Competition
There are some excellent examples of factor theory. Korea has relatively low labor
costs, so it exports labor-intensive goods such as apparel and electronic assemblies.
Sweden’s iron ore is low in impurities, which contributes to a strong Swedish steel
­in­dustry. As compelling as it is, there are counterexamples to the factor endowment
­theory as well. For example, after the Korean War, South Korea developed and excelled
in several highly capital-intensive industries such as steel and shipbuilding even though
the country was cash poor. Also, many countries have similar factor endowments, but
some seem to excel in certain industries, nonetheless. These examples suggest that factor endowments do not explain all cases of nations with dominant industry segments.
Porter (1990) suggests the following four determinants of national advantage:
1. Factor conditions (previously discussed).
2. Demand conditions. If domestic consumers are sophisticated and demanding,
they apply pressure on local industry to innovate faster, which gives firms an edge
­internationally. Consumers of electronics in Japan are very demanding, thus positioning this industry competitively in the international marketplace.
3. Related and supporting industries. Having world-class suppliers nearby is a strong
advantage. For example, the Italian footwear industry is supported by a strong
leather industry and a strong design industry.
4. Firm strategy, structure, and rivalry. The manner in which firms are organized and
managed contributes to their international competitiveness. Japanese management
style is distinctly different from American. In Germany, many senior executives
possess a technical background, producing a strong inclination to product and process improvement. In Italy, there are many small family-owned companies, which
encourages individualism.
Even though Porter makes a very convincing argument for national competitive
­advantage in some industries, there is a debate among economists as to whether the
­notion of international competitiveness makes any sense at all. Companies compete,
not countries. This is the point of view taken by Paul Krugman (1994). According to
Krugman, the United States and Japan are simply not competitors in the same way that
Ford and Toyota are. The standard of living in a country depends on its own ­domestic
economic performance and not on how it performs relative to other countries.
Krugman argues that too much emphasis on international competitiveness can lead to
misguided strategies. Trade wars are much more likely in this case. This was the case in
mid-1995 when the Clinton administration was planning to impose high t­ ariffs on makers
of Japanese luxury cars. Most economists agree that trade wars and their c­ onsequences,
such as tariffs, benefit no one in the long run. Another problem that arises from national
competitive pride is it can lead to poorly conceived government expenditures. France
has spent billions propping up its failing computer industry. (­ Certainly, not all government investments in domestic industry can be considered a mistake. The Japanese government, for example, played a major role in nurturing the flat-panel display industry.
Japanese-based firms now dominate this multibillion dollar industry.)
Another point supporting Krugman’s position is that the lion’s share of U.S. gross
­domestic product (GDP) is consumed in the United States, thus making a firm’s success
in our domestic market more important than its success in the world market. Krugman
agrees that productivity growth is a valid concern. He argues, however, that we should be
more productive in order to produce more, not to better our international competitors.
The debate over competitive advantage will continue. Policy makers need to be
aware of all points of view and weigh each carefully in formulating policy. Although
Krugman makes several telling points, there is no question that globalization is a trend
nah77856_ch01.indd 12
17/12/14 6:02 PM
Snapshot Application
Global Manufacturing Strategies in
the Automobile Industry
Consider the following four foreign automobile manufacturers: Honda, Toyota, BMW, and Mercedes Benz. As
everyone knows, Honda and Toyota are Japanese companies and BMW and Mercedes are German companies.
The four account for the lion’s share of foreign nameplates sold in the U.S. auto market. However, many assume that these cars are manufactured in their home
countries. In fact, depending on the model, it could be
more likely that a consumer buying a Honda, Toyota,
BMW, or Mercedes is buying a car manufactured in the
United States.
Honda was the first of the foreign automakers to commit to a significant investment in U.S.-based manufacturing facilities. Honda’s first U.S. facility was its Marysville
Motorcycle plant, which started production in 1979. Honda
must have been pleased with the Ohio-based f­ acility, since
an automobile plant followed shortly. Automobile production in Marysville began in 1982. Today, Honda operates four plants in west-central Ohio, producing the Accord sedan and coupe, the Acura TL sedan, the Honda Civic
line, and the Honda E
­ lement, with the c­ apacity to produce
a whopping 440,000 vehicles annually.
Next to make a significant commitment in U.S.
production facilities was Toyota. Toyota’s plant in
­
­Georgetown, Kentucky, has been producing automobiles
1.3 Competing in the Global Marketplace 13
since 1986 and accounts for all of the Camrys sold in the
domestic market. It is interesting to note that the Honda
Accord and the Toyota Camry are two of the biggest-­
selling models in the United States, and are also produced
here. They also top almost all reliability surveys.
The two German automakers were slower to commit
to U.S.-based manufacturing facilities. BMW launched
its Spartanburg, South Carolina, plant in March of 1995.
BMW produces both the Z series sports cars and its SUV
line in this plant. It is interesting to note that BMW’s
big sellers, its 3, 5, and 7 series sedans, are still manufactured in Germany.
Mercedes was the last of these four to make a significant commitment to production facilities here. The facility in Tuscaloosa, Alabama, is dedicated to producing the
line of Mercedes SUVs. As with BMW, the more popular
C, E, and S class sedans are still manufactured in Germany.
(One might ask why Volkswagen is not on this list.
In fact, Volkswagen has 45 separate manufacturing facilities located in 18 countries around the world, but
no significant manufacturing presence in the mainland
United States.)
Sources: Honda’s Web site (http://www.ohio.honda.com/),
Toyota’s Web site (http://www.toyota.com), Autointell’s Web
site (http://www.autointell-news.com/european_companies/
BMW/bmw3.htm), Mercedes Benz’s Web site (http://www
.mbusi.com/).
that shows no sign of reversing. We cannot stick our heads in the sand and say that
­foreign markets are not important to us. Economic borders are coming down all across
the globe.
Problems for Sections 1.1–1.3
1.Why is it undesirable for the United States to evolve into a service economy?
2.What disadvantages do you see if the chief executive officer (CEO) is primarily
concerned with short-term ROI?
3. Can you think of companies that have gone out of business because they
­focused only on cost and were not able to achieve a minimum quality s­ tandard?
4. What are the different quality standards referred to in the example comparing the
Honda Civic and the Ferrari?
5. Discuss the pros and cons of trade barriers from the industry point of view and
from the consumer point of view.
6. What are the advantages and disadvantages of producing new products in existing
facilities?
7. What are the four determinants of national advantage suggested by Porter? Give
examples of companies that have thrived as a result of each of these f­ actors.
8. What factor advantage favors the aluminum industry in the United States over
Japan and makes aluminum much cheaper to produce here? (Hint: Aluminum
nah77856_ch01.indd 13
17/12/14 6:02 PM
nah77856
14 Chapter One Strategy and Competition
9.
10.
11.
12.
p­ roduction is very energy intensive. In what part of the country is an inexpensive
­energy source available?)
Paul Krugman argues that because most of our domestic product is consumed
­domestically, we should not dwell on international competition. What ­industries
in the United States have been hardest hit by foreign competition? What are the
potential threats to the United States if these industries fail a­ ltogether?
Krugman points out some misguided government programs that have resulted
from too much emphasis on international competitiveness. What risks are there
from too little emphasis on international competitiveness?
Consider the Snapshot Application in this section concerning foreign automakers
locating manufacturing facilities in the United States. Discuss the advantages and
disadvantages of the strategy of locating manufacturing facilities where the product
is consumed rather than where the company is located.
The North American Free Trade Agreement (NAFTA) was established in 1994
under the Clinton administration.
a) What was the purpose of NAFTA?
b) At the time, political opponents characterized “the big sucking sound” as jobs
would be lost as a result. Is there any evidence that this was, in fact, the case?
1.4 Strategic Initiatives: Reengineering
the Business Process
Seemingly on schedule, every few years a hot new production control method or management technique comes along, almost always described by a three-letter acronym.
While it is easy to be skeptical, by and large, the methods are sound and can have
substantial value to corporations when implemented intelligently. Business ­process
­reengineering (BPR) caught on after the publication of the book by Hammer and
Champy (1993). BPR is not a specific technique, such as materials requirements
planning or a p­ roduction-planning concept like just-in-time. Rather, it is the idea that
entrenched business processes can be changed and can be improved. The process is
one of questioning why things are done a certain way, and not accepting the answer,
­“because that’s the way we do it.’’
Hammer and Champy, who define BPR as “starting over,’’ provide several examples
of successful reengineering efforts. The first is the IBM Credit Corporation, a wholly
owned subsidiary of IBM, that, if independent, would rank among the Fortune 100 ser­
vice companies. This arm of IBM is responsible for advancing credit to new customers
purchasing IBM equipment. The traditional credit approval process f­ ollowed five steps:
1. An IBM salesperson would call in with a request for financing and the request
would be logged on a piece of paper.
2. Someone carried the paper upstairs to the credit department, where someone else
would enter the information into a computer system and check the potential borrower’s credit rating. The specialist wrote the results of the credit check on a piece
of paper, which was sent off to the business practices department.
3. A third person, in the business practices department, modified the standard loan
document in response to the customer request. These modifications, which were
done on yet another computer system, were then attached to the original request
form and the credit department specialist’s report.
nah77856_ch01.indd 14
17/12/14 6:02 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 19
1.5 Strategic Initiatives: Reengineering the Business Process 15
Process 19
1.4 Strategic
4. Next the request went to a pricer, who keyed the information into a spreadsheet to
determine the appropriate interest rate to charge the customer. The pricer’s recommendation was written on a piece of paper and delivered (with the other papers) to
the clerical group.
5. The information was turned into a quote letter that would be delivered to the field
salesperson by Federal Express.
This process required an average of six days and sometimes as long as two weeks.
Sales reps logged endless complaints about this delay: during this time, the customer
could find other financing or another vendor. In an effort to see if this process could be
streamlined, two senior managers decided to walk a new request through all five steps,
asking personnel to put aside what they were doing and process it as they normally
would. They found that the entire five-step process required an average of only 90 minutes of work! The rest of the time, requests were either in transit from one department
to another or queueing up on somebody’s desk waiting to be processed. Clearly the
problem did not lie with the efficiency of the personnel but with the design of the credit
approval process itself.
The solution was simple: the four specialists handling each loan request were replaced by a single loan generalist who handled each request from beginning to end.
Up-to-date software was designed to support the generalist, who had no trouble dealing with most requests. The credit approval process was designed assuming that each
request was sufficiently complex to require someone with special knowledge in each
area. In truth, most requests were routine, and specialists generally did little more than
a simple table lookup to determine the appropriate figure.
What was the result of this change? The six-day turnaround for loan requests was
slashed to only four hours! And this was accomplished with fewer personnel and with
a hundredfold increase in the number of deals handled.
While each reengineering effort requires careful thought and no two solutions will
be exactly alike, Hammer and Champy (1993) suggest that reengineering efforts utilize
the following general principles:
1. Several jobs are combined into one. Few examples of BPR are as dramatic as that
of IBM Credit, but there are other success stories in the literature as well. Many of the
successful cases have a common thread: the reduction of a complex process requiring
many steps to a simpler one requiring fewer steps. In the case of IBM Credit, a five-step
process was reduced to only a single step. This suggests a general principle. The IBM
Credit process was a natural evolution of the concept of division of labor. The economist
Adam Smith espoused this principle as far back as the 18th century (see the quote from
The Wealth of Nations at the beginning of Section 1.10 of this chapter). However, a good
thing can be carried too far. If one divides a process into too many steps, one eventually
reaches the point of diminishing returns. BPR’s most dramatic successes have come from
complex processes that were simplified by reducing the number of steps required.
2. Workers make decisions. One goal is to reduce the number of levels of reporting
by allowing workers to make decisions that were previously reserved for management.
In the case of IBM Credit, most decisions once reserved for specialists are now done
by a single generalist. Giving workers greater decision-making power may pose a
threat to management, who might see such a step as encroaching on their prerogatives.
3. The steps in the process are performed in a natural order. Process steps should not
be performed necessarily in rigid linear sequence, but in an order that makes sense in the
context of the problem being solved. In particular, in many cases, some tasks can be
nah77856_ch01.indd 15
17/12/14 6:02 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 20
nah77856
20
Chapter One
Strategy and Competition
16 Chapter
One Strategy
done simultaneously rather than in sequence. (These ideas, of course, are well known
and form the basis for the concepts of project management in Chapter 10.)
4. Processes should have multiple versions. One should allow for contingencies, not
by designing multiple independent processes, but by designing one flexible process that
can react to different circumstances. In the case of IBM Credit, for example, the final
credit issuance process had three versions: one for straightforward cases (handled by
computer), one for cases of medium difficulty (handled by the deal structurer), and one
for difficult cases (performed by the deal structurer with help from specialist advisers).
5. Work is performed where it makes the most sense. One of the basic principles of
reengineering is not to carry the idea of division of labor too far. Another is not to carry
the idea of centralization too far. For example, in most companies, purchasing is done
centrally. This means that every purchase request is subject to the same minimum overhead in time and paperwork. A consequence might be that the cost of processing a request exceeds the cost of the item being purchased! A great deal can be saved in this
case by allowing individual departments to handle their own purchasing for low-cost
items. (Hammer and Champy discuss such a case.)
The authors list several other basic principles, involving minimizing checks and
reconciliations, having a single point of contact, and being able to employ hybrid
centralized/decentralized operations.
It is easier to list the steps one might consider in a reengineering effort than to actually implement one. In the real world, political realities cannot be ignored. For many of
the success stories in the literature, not only are the processes simplified, but the headcount of personnel is reduced as well. It is certainly understandable for employees to see
BPR as a thinly veiled excuse for downsizing (euphemistically called “right-sizing’’).
This was exactly the case in one financial services company. When word got out that
management was planning a reengineering effort, most assumed that there would
be major layoffs. Some even thought the company was on the verge of bankruptcy.
In another instance, union leadership saw reengineering as a means for management to
throw away the job categories and work rules they had won in hard-fought negotiations
over the years, and persuaded the members to strike. In a third case, a senior manager
was unhappy with the potential loss of authority that might accompany a reengineering
effort. He resigned to start his own company. (These examples are related in a follow-up
book by Hammer and Stanton, 1995.)
These stories show that starting a reengineering effort is not without risks. Rarely is
the process as simple as IBM’s. Reengineering has been described by Ronald Compton, the CEO of Aetna Life and Casualty, as “agonizingly, heartbreakingly, tough.’’
There needs to be some cost–benefit analysis done up front to be sure the potential
gains compensate for the risks.
Process optimization is not new. In its early years, the field of industrial engineering dealt with optimal design of processes, setting standards using time and motion
studies, and flowcharting for understanding the sequence of events and flow of material in a factory. Why is BPR different? For one, BPR is concerned with business
process flows rather than manufacturing process flows. Second, the concept is not one
of optimizing an existing process, but one of rethinking how things should be done
from scratch. As such, it is more revolutionary than evolutionary. It is likely to be more
disruptive but could have larger payoffs. To make BPR work, employees at every level
have to buy into the approach, and top management must champion it. Otherwise, the
reengineering effort could be a costly failure.
nah77856_ch01.indd 16
17/12/14 6:02 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 21
1.5 Strategic Initiatives: Just-in-Time 17
Just-in-Time 21
1.5 Strategic
1.5 STRATEGIC INITIATIVES: JUST-IN-TIME
Just-in-time (JIT) is a manufacturing process on one hand and a broad-based operations strategy on the other. The process elements of JIT will be discussed in detail in
Chapter 8 as part of a complete analysis of push and pull inventory systems.
However, JIT (or lean production, as it is also known) is a philosophy that includes
treatment of inventory in the plant, relationships with suppliers, and distribution
strategies. The core of the philosophy is to eliminate waste. This is accomplished by
efficient scheduling of incoming orders, work-in-process inventories, and finished
goods inventories.
JIT is an outgrowth of the kanban system introduced by Toyota. Kanban is a
Japanese word meaning card or ticket. Originally, kanban cards were the only means of
implementing JIT. The kanban system was introduced by Toyota to reduce excess workin-process (WIP) inventories. Today, JIT is more ambitious. Both quality control systems and relationships with suppliers are part of an integrated JIT system. JIT systems
can be implemented in ways other than using kanban cards. Integrating JIT philosophies with sophisticated information systems makes information transfer faster. The
speed with which information can be transferred from one part of the firm to another is
an important factor in the success of the JIT system.
JIT is a philosophy of operating a company that includes establishing understandings and working relationships with suppliers, providing for careful monitoring of
quality and work flow, and ensuring that products are produced only as they are
needed. Although JIT can be used simply as it was originally designed by Toyota,
namely as a means of moving work-in-process (WIP) from one work center to another,
proponents of the method recommend much more. They would have a firm integrate
the JIT philosophy into its overall business strategy.
Inventory and material flow systems are classified as either push or pull
systems. A push system is one in which decisions concerning how material will
flow through the system are made centrally. Based on these decisions, material is
produced and “pushed’’ to the next level of the system. A typical push system is
materials requirements planning (MRP), which is discussed in detail in Chapter 8.
In MRP, appropriate production amounts for all levels of the production hierarchy
are computed all at once based on forecasts of end-product demand and the relationship between components and end items. In JIT, production is initiated at one
level as a result of a request from a higher level. Units are then “pulled’’ through the
system.
JIT has many advantages over conventional systems. Eliminating WIP inventories
results in reduced holding costs. Less inventory means less money tied up in inventory. JIT also allows quick detection of quality problems. Since units are produced
only as they are needed, the situation in which large amounts of defective WIP
inventory are produced before a quality problem is detected should never occur in
a properly running JIT system. JIT also means that relationships with suppliers
must be tightened up. Suppliers must be willing to absorb some uncertainty and adjust delivery quantities and the timing of deliveries to match the rates of product
flows.
Part of what made the kanban system so effective for Toyota was its success in
reducing setup times for critical operations. The most dramatic example of setup time
reduction is the so-called SMED, or single-minute exchange of dies. Each time a major
change in body style is initiated, it is necessary to change the dies used in the process.
nah77856_ch01.indd 17
17/12/14 6:02 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 22
nah77856
22
Chapter One
Strategy and Competition
18 Chapter
One Strategy
The die-changing operation typically took from four to six hours. During the
die-changing operation the production line was closed down. Toyota management
heard that Mercedes Benz was able to reduce its die-changing operation to less than
one hour. Realizing that even more dramatic reductions were possible, Toyota set about
focusing on the reduction of the time required for die changing. In a series of dramatic
improvements, Toyota eventually reduced this critical operation to only several minutes. The essential idea behind SMED is to make as many changes as possible off-line,
while the production process continues.
An important part of JIT is forming relationships with suppliers. What separates JIT
purchasing from conventional purchasing practices? Freeland (1991) gives a list of
characteristics contrasting the conventional and JIT purchasing behavior. Some of
these include
Conventional Purchasing
1.
2.
3.
4.
5.
6.
Large, infrequent deliveries.
Multiple suppliers for each part.
Short-term purchasing agreements.
Minimal exchange of information.
Prices established by suppliers.
Geographical proximity unimportant.
JIT Purchasing
1.
2.
3.
4.
5.
6.
Small, frequent deliveries.
Few suppliers; single sourcing.
Long-term agreements.
Frequent information exchange.
Prices negotiated.
Geographical proximity important.
In his study, Freeland notes that the industries that seemed to benefit most from
JIT purchasing were those that typically had large inventories. Companies without
JIT purchasing tended to be more job-shop oriented or make-to-order oriented.
Vendors that entered into JIT purchasing agreements tended to carry more safety
stock, suggesting manufacturers are reducing inventories at the expense of the vendors. The JIT deliveries were somewhat more frequent, but the differences were not
as large as one might expect. Geographical separation of vendors and purchasers
was a serious impediment to successful implementation of JIT purchasing. The
automotive industry was one that reported substantial benefit from JIT purchasing
arrangements. In other industries, such as computers, the responses were mixed;
some companies reported substantial benefits and some reported few benefits.
Although reducing excess work-in-process inventory can have many benefits, JIT is
not necessarily the answer for all manufacturing situations. According to Stasey and
McNair (1990),
Inventory in a typical plant is like insurance, insurance that a problem in one area of a plant
won’t affect work performed in another. When problems creating the need for insurance
are solved, then inventories disappear from the plant floor.
The implication is that we merely eliminate all sources of uncertainty in the plant
and the need for inventories disappears. The problem is that there are some sources of
variation that can never be eliminated. One is variation in consumer demand. JIT is effective only if final demand is regular. Another may be sources of variation inherent in
the production process or in the equipment. Can one simply legislate away all sources
of uncertainty in the manufacturing environment? Of course not. Hence, although the
underlying principles of JIT are sound, it is not a cure-all and will not necessarily be
the right method for every production situation.
nah77856_ch01.indd 18
17/12/14 6:02 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 23
1.6 Strategic Initiatives: Time-Based Competition 19
Competition 23
1.6 Strategic
1.6 STRATEGIC INITIATIVES: TIME-BASED COMPETITION
Professor Terry Hill of the London School of Business has proposed an interesting way
to look at competitive factors. He classifies them into two types: “qualifiers’’ and
“order winners.’’ A product not possessing a qualifying factor is eliminated from consideration. The order winner is the factor that determines who gets the sale among the
field of qualifiers.
Two factors about which we hear a great deal are quality and time to market. In the
past decade, the Japanese and Germans gained a loyal following among U.S. consumers
by producing quality products. American firms are catching up on the quality dimension. From the discussion in Section 1.6, we see that successful U.S.–based companies
have been able to produce products that match the defect rates of foreign competitors. If
this trend continues, product quality will be assumed by the consumer. Quality may
become an order qualifier rather than an order winner.
If that is the case, what factors will determine order winners in years to come? Japanese automobile companies provided and continue to provide high-quality automobiles. In recent years, however, the major automobile producers in Japan have begun to
focus on aesthetics and consumer tastes. They have branched out from the stolid small
cars of the 1970s and 1980s to new markets with cars such as the Toyota-made Lexus
luxury line and Mazda’s innovative and successful Miata.
The timely introduction of new features and innovative design will determine the
order winners in the automobile industry. In the computer industry, Compaq built its
reputation partly on its ability to be the first to market with new technology. Time-based
competition is a term that we will hear more and more frequently in coming years.
What is time-based competition? It is not the time and motion studies popular
in the 1930s that formed the basis of the industrial engineering discipline. Rather,
according to Blackburn (1991),
Time-based competitors focus on the bigger picture, on the entire value-delivery system.
They attempt to transform an entire organization into one focused on the total time
required to deliver a product or service. Their goal is not to devise the best way to
perform a task, but to either eliminate the task altogether or perform it in parallel with
other tasks so that over-all system response time is reduced. Becoming a time-based
competitor requires making revolutionary changes in the ways that processes are
organized.
Successful retailers understand time-based competition. The success of the fashion
chains The Gap and The Limited is due largely to their ability to deliver the latest
fashions to the customer in a timely manner. Part of the success of the enormously
successful Wal-Mart chain is its time-management strategy. Each stock item in a WalMart store is replenished twice a week, while the industry average is once every two
weeks. This allows Wal-Mart to achieve better inventory turnover rates than its
competition and respond more quickly to changes in customer demand. Wal-Mart’s
strategies have enabled it to become the industry leader, with a growth rate three
times the industry average and profits two times the industry average (Blackburn,
1991, Chapter 3).
Time-based management is a more complex issue for manufacturers, and in some
industries it is clearly the key factor leading to success or failure. The industry leaders
in the dynamic random access memory (DRAM) industry changed four times between
1978 and 1987. In each case, the firm that was first to market with the next-generation
nah77856_ch01.indd 19
17/12/14 6:02 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 24
nah77856
20 Chapter
One Strategy
24
Chapter One
Strategy and Competition
DRAM dominated that market. The DRAM experience is summarized in the following
table (Davis, 1989):
First Year
of Volume
Production
Product
Firm
Year
Introduced
16 K
Mostek
1976
1978
64 K
Hitachi
1979
1982
256 K
NEC
1982
1984
1 MB
Toshiba
1985
1987
Market Leaders
in First Year of
Volume Production
Mostek (25%)
NEC (20%)
Hitachi (19%)
NEC (15%)
NEC (27%)
Hitachi (24%)
Toshiba (47%)
Mitsubishi (16%)
I am aware of no other example that shows so clearly and so predictably the value of
getting to the market first.
1.7 STRATEGIC INITIATIVES: COMPETING ON QUALITY
What competitive factors do American managers believe will be important in the next
decade? Based on a survey of 217 industry participants, the following factors were
deemed as the most important for gaining a competitive edge in the coming years; they
are listed in the order of importance.
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
Conformance quality
On-time delivery performance
Quality
Product flexibility
After-sale service
Price
Broad line (features)
Distribution
Volume flexibility
Promotion
In this list we see some important themes. Quality and time management emerge as
leading factors. Quality control was brought to public attention with the establishment
of the prestigious Malcolm Baldrige Award (modeled after the Japanese Deming Prize,
which has been around a lot longer). Quality means different things in different contexts, so it is important to understand how it is used in the context of manufactured
goods. A high-quality product is one that performs as it was designed to perform.
Products will perform as they are designed to perform if there is little variation in the
manufacturing process. With this definition of quality, it is possible for a product with a
poor design to be of high quality, just as it is possible for a well-designed product to be
of poor quality. Even granting this somewhat narrow definition of quality, what is the
best measure? Defect rates are a typical barometer. However, a more appropriate measure might be reliability of the product after manufacture. This measure is typically used
to monitor quality of products such as automobiles and consumer electronics.
There has been an enormous groundswell of interest in the quality issue in the United
States in recent years. With the onslaught of Japanese competition, many American
nah77856_ch01.indd 20
17/12/14 6:02 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 25
1.7 Strategic
1.8 Strategic Initiatives: Competing on Quality 21
Quality 25
industries are fighting for their lives. The business of selling quality is at an all-time
high. Consulting companies that specialize in providing quality programs to industry,
such as the Juran Institute and Philip Crosby Associates, are doing a booming business.
The question is whether American firms are merely paying lip service to quality or are
seriously trying to change the way they do business. There is evidence that, in some
cases at least, the latter is true.
For example, in a comparison of American and Japanese auto companies, quality as
measured by defects reported in the first three months of ownership declined significantly from 1987 to 1990 for U.S. companies, narrowing the gap with Japan significantly. The Buick Division of General Motors, a winner of the Baldrige Award, has
made dramatic improvements along these lines. Between 1987 and 1990 Buick decreased this defect rate by about 70 percent, equaling the rate for Hondas in 1990
(Business Week, October 22, 1990).
There are many success stories in U.S. manufacturing. Ford Motors achieved dramatic success with the Taurus. Ford improved both quality and innovation, providing
buyers with reliable and technologically advanced cars. In 1980, James Harbour
reported that Japanese automakers could produce a car for $1,500 less than their
American counterparts. That gap has been narrowed by Ford to within a few
hundred dollars. Part of Ford’s success lies in former CEO Donald Petersen’s
decision not to invest billions in new plants incorporating the latest technology as
GM did in the mid-1980s. This is only part of the story, however. According to Faye
Wills (1990),
If you are looking for surprise answers to Ford’s ascendancy, for hidden secrets, forget it.
Good solid everyday management has turned the trick—textbook planning and execution,
common-sense plant layouts and procedures, intelligent designs that not only sell cars, but
also cut costs and bolster profit margins. It’s that simple.
We can learn from our successes. The machine tool industry was one in which the
Japanese made dramatic inroads in the 1980s. Many American firms fell to the onslaught of Asian competition, but not the Stanley Works of New Britain, Connecticut. In 1982, the firm’s president was considering whether Stanley should remain in
the hardware business as Asian firms flooded the U.S. market with low-priced hammers, screwdrivers, and other tools. Stanley decided to fight back. It modernized its
plants and introduced new quality control systems. Between 1982 and 1988 scrap
rates dropped from 15 percent to only 3 percent at New Britain. Stanley not only
met the competition head-on here at home, but also competed successfully in Asia.
Stanley now runs a profitable operation selling its distinctive yellow tape measures
in Asia.
Where are most PC clones made? Taiwan? Korea? Guess again. The answer may surprise you: Texas. Two Texas firms have been extremely successful in this marketplace.
One is Compaq Computer (now part of HP), which entered the market in the early 1980s
with the first portable PC. It continued to build well-designed and high-quality products, and rose to command 20 percent of the world’s PC market. Compaq established itself as a market leader. The other successful PC maker, Dell Computer, is also from
Texas. The sudden rise of Dell is an interesting story. Michael Dell, a former University
of Texas student, started reselling IBM PCs in the early 1980s. He later formed PC’s
Limited, which marketed one of the first mail-order PC clones. Dell is now a market
leader in the PC marketplace, offering a combination of state-of-the-art designs, highquality products, and excellent service.
Another American firm that has made a serious commitment to quality is Motorola.
Motorola, winner of the Baldrige Award in 1987, has been steadily driving down the
nah77856_ch01.indd 21
17/12/14 6:02 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 26
22 Chapter
One Strategy
26
Chapter One
Strategy and Competition
rate of defects in its manufactured products. Defects were reported to be near 40 parts
per million at the end of 1991, down from 6,000 parts per million in 1986. Motorola
has announced that its goal is to reach six-sigma (meaning six standard deviations
away from the mean of a normal distribution), which translates to 3.4 parts per million.
Motorola feels that the process of applying for the Baldrige Award was so valuable that
it now requires all its suppliers to apply for the award as well.
Success stories like these show that the United States can compete successfully with
Japan and other overseas rivals on the quality dimension. However, total quality management must become ingrained into our culture if we are going to be truly world class.
The fundamentals must be there. The systems must be in place to monitor the traditional quality measures: conformance to specifications and defect-free products. However, quality management must expand beyond statistical measures. Quality must
pervade the way we do business, from quality in design, quality in manufacture, and
quality in building working systems with vendors, to quality in customer service and
satisfaction.
Problems for Sections 1.4–1.7
13. What is an operational definition of quality? Is it possible for a 13-inch TV selling
for $100 to be of superior quality to a 35-inch console selling for $1,800?
14. Studies have shown that the defect rates for many Japanese products are much
lower than for their American-made counterparts. Speculate on the reasons for
these differences.
15. What does “time-based competition” mean? Give an example of a product that
you purchased that was introduced to the marketplace ahead of its competitors.
16. Consider the old maxim, “Build a better mousetrap and the world will beat a path
to your door.’’ Discuss the meaning of this phrase in the context of time-based
competition. In particular, is getting to the market first the only factor in a product’s eventual success?
17. What general features would you look for in a business process that would make that
process a candidate for reengineering? Discuss a situation from your own experience
in which it was clear that the business process could have been improved.
18. In what ways might the following techniques be useful as part of a reengineering
effort?
• Computer-based simulation
• Flowcharting
• Project management techniques
• Mathematical modeling
• Cross-functional teams
19. What problems can you foresee arising in the following situations?
a. Top management is interested in reengineering to cut costs, but the employees
are skeptical.
b. Line workers would like to see a reengineering effort undertaken to give them
more say-so in what goes on, but management is uninterested.
20. Just-in-time has been characterized as a system whose primary goal is to eliminate waste. Discuss how waste can be introduced in (a) relationships with vendors,
nah77856_ch01.indd 22
17/12/14 6:02 PM
1.8 Strategic Initiatives: Servicization 23
(b) receipt of material into the plant, and (c) movement of material through the plant.
How do JIT methods cut down on these forms of waste?
21. In what ways can JIT systems improve product quality?
1.8 Strategic Initiatives: Servicization
Hyundai is a huge multinational company based in South Korea, known as a chaebol,
meaning a collection of diverse companies under a single umbrella. Prior to spinning
off several of its businesses as separate companies following the Asian financial crisis
in 1997, Hyundai was the largest chaebol in South Korea. The Hyundai Motor ­Company
was established in 1967 and first began selling cars in the United States in 1986. At
that time, Japanese, American, and European carmakers were firmly entrenched in
the lucrative U.S. market. One way in which Hyundai sought to differentiate itself
from its competitors was by offering an exceptional warranty. Today Hyundai offers
a ­comprehensive warranty package, including a seven year/100,000 mile powertrain
warranty. Sales of Hyundai models have risen steadily since 1986, with the company
now offers high-end luxury vehicles, along with its low-cost entry models.
While there is no question that competitive pricing and improving reliability and performance of its products account for much of the company’s success, one cannot deny that their
exceptional warranties played a role as well. We will use the term “servicization” to describe
the trend of manufacturing companies to bundle additional services with their products.
Adding services is a means for firms to gain an edge over their competitors, and to provide
an alternative to inexpensive Asian labor. (Note that the Europeans have coined the term
“servitization”. The term appears to have been first used by Vandermerwe and Rada (1988).
Cost-only considerations have driven much of the worldwide manufacturing to
China. The Chinese economy has benefitted enormously from the investments from
foreign economies, led by the United States and Japan. However, there is more to
being an effective manufacturer than low labor rates. The quality of service in many
overseas factories can be disappointing. Quality problems are common and turnaround
times can be crippling. Such factors have lead many firms to rethink their decision to
subcontract their manufacturing to Chinese factories.
Consider the case of Sleek Audio, a producer of high-end earphones. Following the trend
of his competitors, the CEO, Mark Krywko, decided to have his product manufactured in
a factory in Dongguan, China. Unfortunately, Krywko and his son Jason found it necessary to travel to Dongguan every few months because of persistent quality problems. For
example, an entire shipment of 10,000 earphones had to be scrapped because of improper
welding. Delivery delays resulted in emergency air freighting to meet promised deadlines.
Furthermore, design changes took months to be implemented. As problems continued to
mount, the Krywko’s finally decided they’d had enough. Once the decision to come back to
the states was made, Sleek Audio had no problem finding a suitable domestic partner. The
earphones are now manufactured by Dynamic Innovations, whose facility is located only
15 minutes away from the company’s headquarters in Palmetto, Florida. And how has the
decision panned out? After more than a year producing in the United States, Sleek Audio is
projecting 2011 to be the most profitable year in the company’s history (Koerner, 2011).
Servicization: Moving Downstream
Wise and Baumgartner (1999) have noted that many manufacturing companies have
moved their energies downstream to remain competitive. As the installed base of products
increase, and the demand for new products decrease, firms are changing their b­ usiness
model in order to remain competitive. The focus is no longer on the manufacturing
nah77856_ch01.indd 23
17/12/14 6:02 PM
24 Chapter One Strategy and Competition
f­unction alone, but on those services required to operate and maintain products. During
the 40-year period of 1960–2000, the service sector share of the U.S. economy grew
16 percentage points while manufacturing’s share declined by 10 percentage points. What
are some of the downstream activities that manufacturers are becoming more involved
with? The answer is financing, after sales parts and services, and possibly training. The
profit margins on service typically exceed those of manufacturing, thus providing an
­incentive for firms to move in this direction. Of course, the model of vertical integration
is far from new. Part of Henry Ford’s success was the complete vertical integration of the
firm. Ford not only produced the cars, but owned the dealerships where the cars were sold
and even owned the stands of rubber trees used to make tires.
As firms became more specialized, they moved away from vertical integration. However, changing patterns of demands and profits are leading many companies back in this
direction. As an example, the Boeing Company, the world’s foremost manufacturer of
commercial aircraft, has significantly broadened its view of the value chain. The company
now offers financing, local parts supply, ground maintenance, logistics management, and
even pilot training. Servicization can be the key to maintaining competitiveness.
The IBM Story
IBM has become almost synonymous with computing, but few realize that IBM had its
roots in mechanical tabulating machines dating to the late 1800s. The company’s start came
in 1890 when the German immigrant, Herman Hollerith, developed a new process to track
the U.S. census. Hollerith’s concept involved the use of punched cards, which persisted into
the 1960s. The original firm established by Hollerith was the Tabulating Machine Company.
In 1924, 10 years after T.J. Watson joined the firm, the name was changed to International
Business Machines, or IBM. IBM continued to innovate mechanical computing machines,
but was actually a relative late comer into the electronic computer business. In fact, the first
commercial computer was produced by Engineering Research Associates of Minneapolis
in 1950 and sold to the U.S. Navy. Remington Rand produced the Univac one year later,
which was the first commercially viable machine. They sold 46 machines that year at a cost
of over $1 million each. IBM entered the fray in 1953 when it shipped its first computer, the
701. During three years of production, IBM only sold 19 machines. In 1955, AT&T Bell
Laboratories announced the first fully transistorized computer, the TRADIC, and a year
later the Burroughs Corporation threw its hat into the ring and later became a major player
in the computer business. Eventually, Burroughs merged with Sperry Rand to form Unisys.
In 1959, IBM produced its first transistorized-based mainframe computer, the 7000
series. However, it was not until 1964 that the firm became a leader in computer sales.
That was the year that IBM announced the system 360. This was a family of six mutually
compatible computers and 40 peripherals that worked seamlessly together. There is little
question that the system 360 computers were state of the art at the time. Within two years
IBM was shipping 1000 systems per month. IBM has been dominant in the mainframe
business ever since. While all of us are familiar with personal computers of various types
and configurations, the mainframe business has not gone away. Even today, IBM continues to be a presence in the mainframe business with their newest system z architecture.
While the quality of their hardware was an important factor in IBM’s success, it was
not the only factor. What really sealed IBM’s domination of the business market was
the total customer solution. IBM’s industry-specific software and the 360 operating
system were a large part of attracting customers away from its competitors. IBM not
only had one of the most successful sales forces in the business, but also was a master
of after-sales service. Each client would have an SE (systems engineer) assigned to
make sure that their needs were met. As much as anything else, it was IBM’s commitment to after-sales service that locked in their position as market leader in mainframe
nah77856_ch01.indd 24
17/12/14 6:02 PM
1.8 Strategic Initiatives: Servicization 25
computing. Clearly, the idea of edging out competitors by bundling services with products is not new, however, today the servicization concept is becoming an increasingly
important means of gaining competitive advantage. (Information was gathered from
www.computerhistory.org and the IBM company website for this section).
Performance Based Contracts
The impetus behind performance based contracting (PBC) was to reduce excessive
costs in government contracting. Costs were being driven up by unnecessary provisions that specified exactly how each contract was to be carried out. According to Jon
Desenberg of the Washington-based Performance Institute, a think tank dedicated to
improving government performance, the idea “is to let the contracted group come up
with the best possible solution and only pay them based on solving the problem . . . not
on the individual steps and minutia that we have for so many years required.”
Government contracts are a “why,” “how,” and “what” proposition. The “why” is established by the funding agency. Under PBC the “how” is shifted from the government to the
contractor to determine the best way to achieve the “what”. This is not always a win-win for
the contractor. The task of pricing a contract now becomes much more onerous. It can be difficult to estimate all costs in advance, and who pays for contract changes can be problematic.
PBC’s are typical for consumers seeking professional services. A plumber may quote a
fixed price in advance for a simple job, but if there’s uncertainty about the time required,
might want to be compensated on a time plus materials basis. Anyone who has had a
major remodeling of their homes is likely to have entered into a PBC with their contractor.
However, it is rare that the final cost matches the quoted number for a variety of reasons:
weather delays, poor estimation of material costs, difficulty in finding subcontractors, and
more often than not, changes made in the original project by the homeowner.
While PBC’s sound like a good solution for government contracting, it can lead to the
wrong kind of behavior. As an example, Shen (2003) examines the result of PBC’s in the
Maine Addiction Treatment System. Because the system was being measured on the success
of curing addicts, the center had a strong incentive to only treat the less severe cases. This, of
course, runs counter to the purpose of a treatment center; it is the most severe cases of abuse
that need attention. We can conclude that a PBC is not appropriate in all circumstances. For
any contract it is important that incentives be properly aligned with desired outcomes.
Leasing Versus Buying
Leases may be viewed as a service provided to the consumer by the seller. In most
cases, leasing is simply another way for the consumer to finance a purchase. It can
also be viewed as a means for the consumer to reduce risk. The trend towards leasing
of goods and services has increased substantially in recent years for several reasons.
Car leasing has long been an option for consumers. Leases are more popular when
­financing is difficult or expensive to obtain. The prospect of a low monthly payment a­ ttracts
consumers, who may ultimately pay more in the long run. Consider the buyer that likes to
drive a relatively new car and trades in their automobile every three years. This buyer faces
the risk of not being able to accurately predict the trade-in value of the car three years down
the road. In a lease situation, this residual risk is assumed by the seller. The terms of the lease
are predicated on an assumption about the residual value of the car at the end of the lease.
If the manufacturer overestimates the residual value of the car, the consumer benefits by
simply turning the car in at the end of the lease period. If the ­manufacturer u­ nderestimates
the residual value, the leaser wins by purchasing the car at the end of the lease and selling
it for a higher market price. Hence, the manufacturer absorbs the risk of estimating the depreciation, which provides an incentive to the consumer to lease rather than buy. Of course,
since automobiles tend to depreciate most in the first several years, buying and holding a
nah77856_ch01.indd 25
17/12/14 6:02 PM
26 Chapter One Strategy and Competition
car will be a less expensive alternative in the long run for most vehicles—­especially those
that are more durable and more reliable. Since auto leases are typically two to four years,
the buyer that keeps cars for a long time will not choose to lease.
Leasing (that is, renting) versus buying is also an important choice for the consumer when
it comes to choosing how and where to live. While owning a home has been touted as the
“American Dream,” there are clearly many who should not be homeowners. In the United
States, most homeowners have a mortgage on their primary property. A mortgage is simply a
loan provided to homeowner, with the home itself as the collateral. In periods when housing
prices are rising faster than inflation, homeowners have done very well. However, as we’ve
seen in recent years, rising housing prices is not a certainty. There have been several periods
in which housing prices have dropped precipitously, including the Great Depression of the
1930s. In fact, as of this writing, the housing industry has been in a slump since 2008.
The mortgage crisis of 2008 led to a total collapse of many major financial institutions, and nearly caused a worldwide depression. What precipitated this crisis? In a
nutshell, it was awarding mortgages to individuals that did not qualify for them. Why
did this happen? There were a variety of factors including government deregulation,
unscrupulous mortgage brokers, and naive consumers. The stringent standards that
banks traditionally applied before awarding mortgages were thrown out the window.
When real estate prices were on an apparent never-ending upward spiral, homeowners
borrowed well beyond their means to finance other houses, or large capital expenditures such as boats or cars. People who had never owned a home in their lives were
cajoled by unscrupulous mortgage brokers into taking loans they had no chance of
paying off. Housing prices climbed far beyond reasonable levels, and when the bubble
burst, millions of homeowners found themselves underwater (meaning their houses
were worth less than the balance of their mortgage loans). Since many loans were
granted with little or no down payment, many just walked away from their homes
leaving entire neighborhoods vacant. Homes were looted for valuable materials, such
as copper piping, thus assuring that these homes would never be sellable.
Renting is a sensible choice for many. While there appears to be a stigma associated
with being a renter, it became clear from the subprime loan debacle that, in fact, many
folks have no business owning a home. Landlords must absorb the risks of repairs and
price fluctuations. It is no longer obvious that real estate is necessarily a safe investment.
Perhaps the mortgage crises can be viewed as a case of servicization gone out of control.
Green Leasing
A recent trend has been towards green leasing. According to the U.S. Green Building
Council, buildings account for more than one-third of all energy use, carbon dioxide emissions, waste output, and use of raw materials in the United States. While green leasing is
a relatively new concept to American companies, it has been a practice in other parts of
the world for years. Several foreign governments have promulgated environmental-based
rules for their properties. In America, however, the movement is far less centralized.
Most green leasing initiatives in the United States have been proposed by state and
local governments. Government agencies and academic institutions have been the front
runners in green building technology, representing approximately 26 percent of all LEED
certified buildings. LEED, developed by the Green Building Counsel, is an internationally recognized certification system that measures building performance in categories
such as: energy savings, water efficiency, carbon dioxide emissions, indoor environmental quality, and resource management. LEED also provides a framework for stages
throughout a building’s lifecycle, from design and construction to operation and management. LEED provides several levels of certification. Most green building initiatives
provide either incentives or penalties based upon a building’s LEED certification level.
nah77856_ch01.indd 26
17/12/14 6:02 PM
1.9 Matching Process and Product Life Cycles 27
Problems for Section 1.8
22. Define “servicization” and provide an example from your own experience of a
case where services were the deciding factor in a purchase.
23. What are some of the services that IBM provided for its mainframe customers during its meteoric rise in sales in the 1960s?
24. Why can car leasing be viewed as a service? What are the advantages and disadvantages of car leasing from the buyer’s point of view? Why do manufacturers offer leases?
1.9 Matching Process and Product Life Cycles
The Product Life Cycle
The demand for new products typically undergoes cycles that can be identified
and mapped over time. Understanding the nature of this evolution helps to identify
­appropriate strategies for production and operations at the various stages of the ­product
cycle. A typical product life cycle is pictured in Figure 1–3. The product life cycle
consists of four major segments:
1.
2.
3.
4.
Start-up
Rapid growth
Maturation
Stabilization or decline
During the start-up phase, the market for the product is developing, production and
distribution costs are high, and competition is generally not a problem. During this
phase the primary strategy concern is to apply the experiences of the marketplace and
of manufacturing to improve the production and marketing functions. At this time,
­serious design flaws should be revealed and corrected.
Figure 1–3
Annual sales volume
The product life-cycle
curve
Start-up
Rapid growth
Maturation
Stabilization
or decline
Time
nah77856_ch01.indd 27
17/12/14 6:02 PM
nah77856
28 Chapter One Strategy and Competition
The period of rapid growth sees the beginning of competition. The primary strategic
goal during this period is to establish the product as firmly as possible in the market­
place. To do this, management should consider alternative pricing patterns that suit
the various customer classes and should reinforce brand preference among suppliers and customers. The manufacturing process should be undergoing improvements
and standardization as product volume increases. Flexibility and modularization of the
manufacturing function are highly desirable at this stage.
During the maturation phase of the product life cycle, the objective should be to maintain and improve the brand loyalty that the firm cultivated in the growth phase. Management should seek to increase market share through competitive pricing. Cost savings
should be realized through improved production control and product distribution. During
this phase the firm must listen to the messages of the marketplace. Most problems with
product design and quality should have been corrected during the start-up and growth
phases, but additional improvements should also be considered during this phase.
The appropriate shape of the life-cycle curve in the final stage depends on the nature of the product. Many products will continue to sell, with the potential for annual growth continuing almost indefinitely. Examples of such products are commodities such as household goods, processed food, and automobiles. For such products
the company’s primary goals in this phase would be essentially the same as those
­described previously for the maturation phase. Other products will experience a natural ­decline in sales volume as the market for the product becomes saturated or as the
­product becomes obsolete. If this is the case, the company should adopt a strategy of
squeezing out the most from the product or product line while minimizing investment
in new manufacturing technology and media advertising.
Although a useful concept, the product life-cycle curve is not accurate in all circumstances. Marketing departments that base their strategies on the life-cycle curve may
make poor decisions. Dhalla and Yuspeh (1976) report an example of a firm that shifted
advertising dollars from a successful stable product to a new product. The assumption
was that the new product was entering the growth phase of its life cycle and the stable
product was entering the declining phase of its life cycle. However, the new product
never gained consumer acceptance, and because of a drop in the advertising budget, the
sales of the stable product went into a decline and never recovered. They suggest that in
some circumstances it is more effective to build a model that is consistent with the product’s history and with consumer behavior than to blindly assume that all products follow
the same pattern of growth and decline. Although we believe that the life-cycle concept is
a useful way of looking at customer demand patterns in general, a carefully constructed
model for each product will ultimately be a far more e­ ffective planning tool.
The Process Life Cycle
Abernathy and Townsend (1975) have classified three major stages of the manufacturing process life cycle: early, middle, and mature. These phases do not necessarily
coincide ­exactly with the stages of the product life cycle, but they do provide a conceptual framework for planning improvements in the manufacturing process as the
product matures.
In the first phase of the process life cycle, the manufacturing function has the characteristics of a job shop. It must cope with a varied mix of relatively low-volume orders
and be responsive to changes in the product design. The types and quality of the inputs
may vary considerably, and the firm has little control over suppliers.
In the middle phase of the process life cycle, automation begins to play a greater role.
The firm should be able to exert more control over suppliers as the volume of p­ roduction
nah77856_ch01.indd 28
17/12/14 6:02 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 29
1.9 Matching
1.9 Matching Process and Product Life Cycles 29
Cycles 29
The process life cycle
and the experience
curve
Manufacturing cost per unit
FIGURE 1–4
Start-up
Rapid growth
Maturation
Stabilization
or decline
Time
increases. Unit production costs decline as a result of learning effects. The production process may involve batch processing and some transfer lines (assembly lines).
In the last phase of the process life cycle, most of the major operations are automated, the production process is standardized, and few manufacturing innovations are
introduced. The production process may assume the characteristics of a continuous
flow operation.
This particular evolutionary scenario is not appropriate for all new manufacturing
ventures. Companies that thrive on small one-of-a-kind orders will maintain the
characteristics of a job shop, for example. The process life-cycle concept applies to
new products that eventually mature into high-volume items. The issue of matching the
characteristics of the product with the characteristics of the process is discussed
subsequently.
Experience curves show that unit production costs decline as the cumulative number of units produced increases. One may think of the experience curve in terms of the
process life cycle shown in Figure 1–4. An accurate understanding of the relationship
between the experience curve and the process life cycle can be very valuable. By
matching the decline in unit cost with the various stages of the process life cycle, management can gain insight into the consequences of moving from one phase of the
process life cycle into another. This insight will assist management in determining the
proper timing of improvements in the manufacturing process.
The Product–Process Matrix
Hayes and Wheelwright (1979) consider linking the product and process life cycles
using the product–process matrix pictured in Figure 1–5. The matrix is based on four
phases in the evolution of the manufacturing process: (1) jumbled flow, (2) disconnected line flow, (3) connected line flow, and (4) continuous flow. This matrix may be
viewed in two ways. One is to match the appropriate industry in its mature phase with
nah77856_ch01.indd 29
17/12/14 6:02 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 30
nah77856
30 Chapter One Strategy and Competition
FIGURE 1–5
I
The product–process
matrix
Source: Robert H. Hayes and
Steven C. Wheelwright, “Link
Manufacturing Process and
Product Life Cycles” in the
Harvard Business Review
(January–February 1979).
© 1979 by the President and
Fellows of Harvard College; all
rights reserved. Reprinted by
permission.
II
Process structure Low volume, low
Process life-cycle standardization, Multiple products,
one of a kind
low volume
stage
I
Jumbled
flow
(job shop)
IV
Few major
products,
higher volume
High volume, high
standardization,
commodity
products
Commercial
printer
II
Disconnected
line flow
(batch)
Void
Heavy
equipment
III
Connected
line flow
(assembly line)
IV
Continuous
flow
III
Auto
assembly
Void
Sugar
refinery
the appropriate process. This point of view recognizes that not all industries
necessarily follow the process evolution described in the previous section on the
process life cycle. Certain companies or certain products could remain in an early
phase of the process life cycle indefinitely. However, even firms that do not evolve to
a position in the lower right-hand corner of the matrix should, in most cases, be located
somewhere on the diagonal of the matrix.
Located in the upper left-hand corner of this matrix are companies that specialize in
“one of a kind’’ jobs in which the manufacturing function has the characteristics of a
jumbled flow shop. A commercial printer is an example of a jumbled flow shop.
Production is in relatively small lots, and the shop is organized for maximum flexibility.
Farther down the diagonal are firms that still require a great deal of flexibility but
produce a limited line of standardized items. Manufacturers of heavy equipment would
fall into this category because they would produce in somewhat higher volumes. A
disconnected line would provide enough flexibility to meet custom orders while still
retaining economies of limited standardization.
The third category down the diagonal includes firms that produce a line of standard
products for a large-volume market. Typical examples are producers of home
appliances or electronic equipment, and automobile manufacturers. The assembly line
or transfer line would be an appropriate process technology in this case.
nah77856_ch01.indd 30
17/12/14 6:02 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 31
1.10 Learning
1.10 Learning and Experience Curves 31
Curves 31
Finally, the lower right-hand portion of the matrix would be appropriate for products
involving continuous flow. Chemical processing, gasoline and oil refining, and sugar
refining are examples. Such processes are characterized by low unit costs, standardization of the product, high sales volume, and extreme inflexibility of the production
process.
What is the point of this particular classification scheme? It provides a means of assessing whether a firm is operating in the proper portion of the matrix; that is, if the
process is properly matched with the product structure. Firms choosing to operate off
the diagonal should have a clear understanding of the reasons for doing so. One example of a successful firm that operates off the diagonal is Rolls-Royce. Another is a
company producing handmade furniture. The manufacturing process in these cases
would have the characteristics of a jumbled flow shop, but competitors might typically
be located in the second or third position on the diagonal.
There is another way to look at the product–process matrix. It can be used to
identify the proper match of the production process with the phases of the product life
cycle. In the start-up phase of product development, the firm would typically be
positioned in the upper left-hand corner of the matrix. As the market for the product
matures, the firm would move down the diagonal to achieve economies of scale.
Finally, the firm would settle at the position on the matrix that would be appropriate
based on the characteristics of the product.
Problems for Section 1.9
25. a. What are the four phases of the manufacturing process that appear in the
product–process matrix?
b. Discuss the disadvantages of operating off the diagonal of the matrix.
26. Give an example of a product that has undergone the four phases of the product
life cycle and has achieved stability.
27. Discuss the following: “All firms should evolve along the diagonal of the
product–process matrix.’’
28. Locate the following operations in the appropriate position on the product–process
matrix.
a. A small shop that repairs musical instruments.
b. An oil refinery.
c. A manufacturer of office furniture.
d. A manufacturer of major household appliances such as washers, dryers, and
refrigerators.
e. A manufacturing firm in the start-up phase.
1.10
LEARNING AND EXPERIENCE CURVES
As experience is gained with the production of a particular product, either by a single
worker or by an industry as a whole, the production process becomes more efficient.
As noted by the economist Adam Smith as far back as the 18th century in his landmark
work, The Wealth of Nations:
The division of labor, by reducing every man’s business to some one simple operation, and
by making this operation the sole employment of his life, necessarily increases very much
the dexterity of the worker.
nah77856_ch01.indd 31
17/12/14 6:02 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
nah77856
Page 32
32 Chapter One Strategy and Competition
By quantifying the relationship that describes the gain in efficiency as the cumulative number of units produced increases, management can accurately predict the eventual capacity of existing facilities and the unit costs of production. Today we recognize
that many other factors besides the improving skill of the individual worker contribute
to this effect. Some of these factors include the following:
•
•
•
•
•
Improvements in production methods.
Improvements in the reliability and efficiency of the tools and machines used.
Better product design.
Improved production scheduling and inventory control.
Better organization of the workplace.
Studies of the aircraft industry undertaken during the 1920s showed that the directlabor hours required to produce a unit of output declined as the cumulative number of
units produced increased. The term learning curve was adopted to explain this phenomenon. Similarly, it has been observed in many industries that marginal production
costs also decline as the cumulative number of units produced increases. The term
experience curve has been used to describe this second phenomenon.
Learning Curves
As workers gain more experience with the requirements of a particular process, or as
the process is improved over time, the number of hours required to produce an additional unit declines. The learning curve, which models this relationship, is also a means
of describing dynamic economies of scale. Experience has shown that these curves are
accurately represented by an exponential relationship. Let Y(u) be the number of labor
hours required to produce the uth unit. Then the learning curve is of the form
Y(u) aub,
where a is the number of hours required to produce the first unit and b measures the
rate at which the marginal production hours decline as the cumulative number of units
produced increases. Traditionally, learning curves are described by the percentage
decline of the labor hours required to produce item 2n compared to the labor hours
required to produce item n, and it is assumed that this percentage is independent of n.
That is, an 80 percent learning curve means that the time required to produce unit 2n is
80 percent of the time required to produce unit n for any value of n. For an 80 percent
learning curve
Y(2u) a(2u) b
2b .80.
b
Y(u)
au
It follows that
b ln(2) ln(.8)
or b ln(.8)ln(2) .3219. (ln is the natural logarithm.)
More generally, if the learning curve is a 100L percent learning curve, then
b ln(L)ln(2).
Figure 1–6 shows an 80 percent learning curve. When graphed on double-log paper,
the learning curve should be a straight line if the exponential relationship we have
assumed is accurate. If logarithms of both sides of the expression for Y(u) are taken, a
linear relationship results, since
ln(Y(u)) ln(a) b ln(u).
nah77856_ch01.indd 32
17/12/14 6:03 PM
nah77856_ch01.qxd
1/25/08
4:03 PM
Page 33
1.10
An 80 percent learning
curve
Y(u)
FIGURE 1–6
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
0
0
10
20
30
40
50
60
Learning and Experience Curves 33
70
80
90
100
u
Linear regression is used to fit the values of a and b to actual data after the logarithm
transformation has been made. (General equations for finding least squares estimators
in linear regression appear in Appendix 2–B.)
Example 1.1
XYZ has kept careful records of the average number of labor hours required to produce one of
its new products, a pressure transducer used in automobile fuel systems. These records are represented in the following table.
Cumulative Number of
Units Produced (A)
10.00
25.00
100.00
250.00
500.00
1,000.00
5,000.00
10,000.00
Ln
(Column A)
Hours Required for
Next Unit (B)
2.30
3.22
4.61
5.52
6.21
6.91
8.52
9.21
9.22
4.85
3.80
2.44
1.70
1.03
0.60
0.50
Ln
(Column B)
2.22
1.58
1.34
0.89
0.53
0.53
0.51
0.69
According to the theory, there should be a straight-line relationship between the logarithm of
the cumulative number of units produced and the logarithm of the hours required for the last unit
of production. The graph of the logarithms of these quantities for the data above appears in
Figure 1–7. The figure suggests that the exponential learning curve is fairly accurate in this case.
Using the methods outlined in Appendix 2–B, we have obtained estimators for the slope and
the intercept of the least squares fit of the data in Figure 1–7. The values of the least squares
estimators are
Intercept 3.1301,
Slope .42276.
Since the intercept is ln(a), the value of a is exp(3.1301) 22.88. Hence, it should have
taken about 23 hours to produce the first unit. The slope term is the constant b. From the
nah77856_ch01.indd 33
17/12/14 6:03 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 34
nah77856
34 Chapter One Strategy and Competition
Log–log plot of XYZ
data
Natural log (time to produce next unit)
FIGURE 1–7
2.4
2.2
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
–0.2
–0.4
–0.6
–0.8
0
2
4
6
8
Natural log (cumulative units produced)
10
equation for b on page 33 we have that
ln(L) b ln(2) (.42276)(.6931) .293.
It follows that L exp(.293) .746.
Hence, these data show that the learning effect for the production of the transducers can
be accurately described by a 75 percent learning curve. This curve can be used to predict the
number of labor hours that will be required for continued production of these particular transducers. For example, substituting u 50,000 into the relationship
Y(u) 22.88u.42276
gives a value of Y(50,000) .236 hour. One must interpret such results with caution, however.
A learning curve relationship may not be valid indefinitely. Eventually the product will reach the
end of its natural life cycle, which could occur before 50,000 units have been produced in this
example. Alternatively, there could be some absolute limit on the number of labor hours required to produce one unit that, because of the nature of the manufacturing process, can never
be improved. Even with these limitations in mind, learning curves can be a valuable planning
tool when properly used.
Experience Curves
Learning curves are a means of calibrating the decline in marginal labor hours as workers become more familiar with a particular task or as greater efficiency is introduced
into the production process. Experience curves measure the effect that accumulated experience with production of a product or family of products has on overall cost and
price. Experience curves are most valuable in industries that are undergoing major
changes, such as the microelectronics industry, rather than very mature industries in
which most radical changes have already been made, such as the automobile industry.
The steady decline in the prices of integrated circuits (ICs) is a classic example of an
experience curve. Figure 1–8 (Noyce, 1977) shows the average price per unit as a function of the industry’s accumulated experience, in millions of units of production, during the period 1964 to 1972. This graph is shown on log–log scale and the points fall
very close to a straight line. This case represents a 72 percent experience curve. That
nah77856_ch01.indd 34
17/12/14 6:03 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 35
1.10
Learning and Experience Curves 35
Prices of integrated
circuits during the
period 1964–1972
Source: Robert N. Noyce,
“Microelectronics,” in Scientific
American, September 1977.
© 1977 by Scientific American,
Inc. All rights reserved.
Reprinted with permission
of the publisher.
Average price per unit (constant dollars)
FIGURE 1–8
1964
20
1966
1968
1972
10
5
2
1
.5
.2
.1
1
2
5
10
20
50
100
200
500 1,000 2,000
Industry’s accumulated experience (millions of units)
“Kinked” experience
curve due to umbrella
pricing
Log of the price per unit
FIGURE 1–9
Log of the cumulative number of units produced
is, the average price per unit declines to about 72 percent of its previous value for each
doubling of the cumulative production of ICs throughout the industry.
Experience curves are generally measured in terms of cost per unit of production. In
most circumstances, the price of a product or family of products closely tracks the cost of
production. However, in some cases umbrella pricing occurs.That is, prices remain fairly
stable during a period in which production costs decline. Later, as competitive pressures
of the marketplace take hold, prices decline more rapidly than costs until they catch up.
This can cause a kink in the experience curve when price rather than cost is measured
against cumulative volume. This type of phenomenon is pictured in Figure 1–9. Hayes
nah77856_ch01.indd 35
17/12/14 6:03 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 36
nah77856
36 Chapter One Strategy and Competition
and Wheelwright (1984, p. 243) give two examples of umbrella pricing and its effect on
the experience curve. The experience curves for the pricing of free-standing gas ranges
and polyvinyl chloride are examples of this phenomenon.
Learning curves have been the subject of criticism in the literature recently on a number of grounds: (a) they lack theoretical justification; (b) they confuse the effects of
learning, economies of scale, and other technological improvements; and (c) they
focus on cost rather than profit (Devinney, 1987). However, it is clear that such curves
are accurate descriptors of the way that marginal labor hours and costs decline as a
function of the cumulative experience gained by the firm or industry.
Learning and Experience Curves and Manufacturing Strategy
We define a learning curve strategy as one in which the primary goal is to reduce costs
of production along the lines predicted by the learning curve. Ford Motors adopted a
learning curve strategy in seeking cost reductions in the Model T during the period 1909
to 1923. Abernathy and Wayne (1974) showed that the selling price of the Model T
during this period closely followed an 85 percent experience curve. Ford’s strategy during this time was clearly aimed at cost cutting; the firm acquired or built new facilities
including blast furnaces, logging operations and saw mills, a railroad, weaving mills,
coke ovens, a paper mill, a glass plant, and a cement plant. This allowed Ford to vertically integrate operations, resulting in reduced throughput time and inventory levels—a
strategy similar in spirit to the just-in-time philosophy discussed earlier in this chapter.
A learning curve strategy may not necessarily be the best choice over long planning
horizons. Abernathy and Wayne (1974) make the argument that when manufacturing
strategy is based on cost reduction, innovation is stifled. As consumer tastes changed
in the 1920s, Ford’s attention to cost cutting and standardization of the Model T manufacturing process resulted in its being slow to adapt to changing patterns of customer
preferences. Ford’s loss was General Motors’s gain. GM was quick to respond to customer needs, recognizing that the open car design of the Model T would soon become
obsolete. Ford thus found itself fighting for survival in the 1930s after having enjoyed
almost complete domination of the market. Survival meant a break from the earlier
rigid learning curve strategy to one based on innovation.
Another example of a firm that suffered from a learning curve strategy is Douglas
Aircraft. The learning curve concept was deeply rooted in the airframe industry.
Douglas made several commitments in the 1960s for delivery of jet aircraft based on extrapolation of costs down the learning curve. However, because of unforeseen changes
in the product design, the costs were higher than anticipated, and commitments for
delivery times could not be met. Douglas was forced into a merger with McDonnell
Company as a result of the financial problems it experienced.
We are not implying by these examples that a learning curve strategy is wrong. Standardization and cost reduction based on volume production have been the keys to
success for many companies. Failure to achieve quick time-to-volume can spell disaster in a highly competitive marketplace. What we are saying is that the learning curve
strategy must be balanced with sufficient flexibility to respond to changes in the
marketplace. Standardization must not stifle innovation and flexibility.
Problems for Section 1.10
29. What are the factors that contribute to the learning curve/experience curve
phenomenon?
30. What is a “learning curve strategy’’? Describe how this strategy led to Ford’s
success up until the mid-1920s and Ford’s problems after that time.
nah77856_ch01.indd 36
17/12/14 6:03 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 37
1.10
Learning and Experience Curves 37
31. What are some of the pitfalls that can occur when using learning curves and
experience curves to predict costs? Refer to the experience of Douglas Aircraft.
32. Consider the example of XYZ Corporation presented in this section. If the learning
curve remains accurate, how long will it take to produce the 100,000th unit?
33. A start-up firm has kept careful records of the time required to manufacture its
product, a shutoff valve used in gasoline pipelines.
Cumulative Number
of Units Produced
50
100
400
600
1,000
10,000
Number of Hours Required
for Next Unit
3.3
2.2
1.0
0.8
0.5
0.2
a. Compute the logarithms of the numbers in each column. (Use natural logs.)
b. Graph the ln(hours) against the ln(cumulative units) and eyeball a straight-line
fit of the data. Using your approximate fit, estimate a and b.
c. Using the results of part (b), estimate the time required to produce the first unit
and the appropriate percentage learning curve that fits these data.
d. Repeat parts (b) and (c), but use an exact least squares fit of the logarithms
computed in part (a).
34. Consider the learning curve derived in Problem 30. How much time will be required
to produce the 100,000th unit, assuming the learning curve remains accurate?
35. Consider the experience curve plotted in Figure 1–10. What percentage experience
curve does this represent?
FIGURE 1–10
Marginal production cost
(for Problem 32)
30
28
26
24
22
20
18
16
14
12
10
8
6
4
2
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Cumulative number of units produced
nah77856_ch01.indd 37
17/12/14 6:03 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 38
nah77856
38 Chapter One Strategy and Competition
36. Discuss the limitations of learning and experience curves.
37. An analyst predicts that an 80 percent experience curve should be an accurate predictor of the cost of producing a new product. Suppose that the cost of the first unit
is $1,000. What would the analyst predict is the cost of producing the
a. 100th unit?
b. 10,000th unit?
1.11 CAPACITY GROWTH PLANNING: A LONG-TERM
STRATEGIC PROBLEM
The capacity of a plant is the number of units that the plant can produce in a given time.
Capacity policy plays a key role in determining the firm’s competitive position in the
marketplace. A capacity strategy must take into account a variety of factors, including
•
•
•
•
Predicted patterns of demand.
Costs of constructing and operating new facilities.
New process technology.
Competitors’ strategy.
Capacity planning is an extremely complex issue. Each time a company considers
expanding existing productive capacity, it must sift through a myriad of possibilities.
First, the decision must be made whether to increase capacity by modifying existing facilities. From an overhead point of view, this is an attractive alternative. It is cheaper to
effect major changes in existing processes and plants than to construct new facilities.
However, such a strategy ultimately could be penny wise and pound foolish. There is
substantial evidence that plants that have focus are the most productive. Diminishing
returns quickly set in if the firm tries to push the productive capacity of a single location beyond its optimal value.
Given the decision to go ahead with construction of a new plant, many issues remain
to be resolved. These include
1. When. The timing of construction of new facilities is an important consideration.
Lead times for construction and changing patterns of demand are two factors that
affect timing.
2. Where. Locating new facilities is a complex issue. Consideration of the logistics
of material flows suggests that new facilities be located near suppliers of raw materials
and market outlets. If labor costs were the key issue, overseas locations might be preferred. Tax incentives are sometimes given by states and municipalities trying to attract
new industry. Cost of living and geographical desirability are factors that would affect
the company’s ability to hire and keep qualified employees.
3. How much. Once management has decided when and where to add new capacity,
it must decide on the size of the new facility. Adding too much capacity means that the
capacity will be underutilized. This is an especially serious problem when capital is
scarce. On the other hand, adding too little capacity means that the firm will soon be
faced with the problem of increasing capacity again.
Economies of Scale and Economies of Scope
Economies of scale are generally considered the primary advantages of expanding
existing capacity. Panzer and Willig (1981) introduced the concept of economies of
scope, which they defined as the cost savings realized from combining the production
nah77856_ch01.indd 38
17/12/14 6:03 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 39
1.11
Capacity Growth Planning: A Long-Term Strategic Problem 39
of two or more product lines at a single location. The idea is that the manufacturing
processes for these product lines may share some of the same equipment and personnel so that the cost of production at one location could be less than at two or more different locations.
The notion of economies of scope extends beyond the direct cost savings that the
firm can realize by combining the production of two or more products at a single
location. It is often necessary to duplicate a variety of support functions at different
locations. These functions include information storage and retrieval systems and clerical and support staff. Such activities are easier to coordinate if they reside at the same
location. The firm also can realize economies of scope by locating different facilities
in the same geographic region. In this way employees can, if necessary, call upon the
talents of key personnel at a nearby location.
Goldhar and Jelinek (1983) argue that considerations of economies of scope support
investment in new manufacturing technology. Flexible manufacturing systems and
computer-integrated manufacturing result in “efficiencies wrought by variety, not
volume.’’ These types of systems, argue the authors, allow the firm to produce multiple
products in small lot sizes more cheaply using the same multipurpose equipment. (Flexible manufacturing systems are discussed in greater detail in Chapter 11.)
Management must weigh the benefits that the firm might realize by combining
product lines at a single location against the disadvantages of lack of focus discussed
previously. Too many product lines produced at the same facility could cause the various manufacturing operations to interfere with each other. The proper sizing and diversity of the functions of a single plant must be balanced so that the firm can realize
economies of scope without allowing the plant to lose its essential focus.
Make or Buy: A Prototype Capacity Expansion Problem
A classic problem faced by the firm is known as the make-or-buy decision. The firm
can purchase the product from an outside source for c1 per unit, but can produce it
internally for a lower unit price, c2 c1. However, in order to produce the product internally, the company must invest $K to expand production capacity. Which strategy
should the firm adopt?
The make-or-buy problem contains many of the elements of the general capacity expansion problem. It clarifies the essential trade-off of investment and economies
of scale. The total cost of the firm to produce x units is K c 2 x. This is equivalent to
Kx c2 per unit. As x increases, the cost per unit of production decreases, since Kx
is a decreasing function of x. The cost to purchase outside is c1 per unit, independent
of the quantity ordered. By graphing the total costs of both internal production and
external purchasing, we can find the point at which the costs are equal. This is known
as the break-even quantity. The break-even curves are pictured in Figure 1–11.
The break-even quantity solves
K c 2 x c1x,
giving x K(c1 c2).
Example 1.2
nah77856_ch01.indd 39
A large international computer manufacturer is designing a new model of personal computer
and must decide whether to produce the keyboards internally or to purchase them from an
outside supplier. The supplier is willing to sell the keyboards for $50 each, but the manufacturer
estimates that the firm can produce the keyboards for $35 each. Management estimates that
expanding the current plant and purchasing the necessary equipment to make the keyboards
would cost $8 million. Should they undertake the expansion?
17/12/14 6:03 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 40
nah77856
40 Chapter One Strategy and Competition
FIGURE 1–11
Break-even curves
Cost to buy = c1x
Cost
Cost to make = K + c2x
K
Break-even
quantity
Number of units (x)
The break-even quantity is
x 8,000,000(50 35) 533,333.
Hence, the firm would have to sell at least 533,333 keyboards in order to justify the $8 million
investment required for the expansion.
Break-even curves such as this are useful for getting a quick ballpark estimate of the
desirability of a capacity addition. Their primary limitation is that they are static. They do
not consider the dynamic aspects of the capacity problem, which cannot be ignored in
most cases. These include changes in the anticipated pattern of demand and considerations of the time value of money. Even as static models, break-even curves are only rough
approximations. They ignore the learning effects of production; that is, the marginal production cost should decrease as the number of units produced increases. (Learning
curves are discussed in detail in Section 1.10.) Depending on the structure of the production function, it may be economical to produce some units internally and purchase
some units outside. Manne (1967) discusses the implications of some of these issues.
Dynamic Capacity Expansion Policy
Capacity decisions must be made in a dynamic environment. In particular, the dynamics of the changing demand pattern determine when the firm should invest in new capacity. Two competing objectives in capacity planning are
1. Maximizing market share
2. Maximizing capacity utilization
A firm that bases its long-term strategy on maximization of capacity utilization runs
the risk of incurring shortages in periods of higher-than-anticipated demand. An alternative strategy to increasing productive capacity is to produce to inventory and let the
inventory absorb demand fluctuations. However, this can be very risky. Inventories can
become obsolete, and holding costs can become a financial burden.
Alternatively, a firm may assume the strategy of maintaining a “capacity cushion.’’
This capacity cushion is excess capacity that the firm can use to respond to sudden demand surges; it puts the firm in a position to capture a larger portion of the marketplace
if the opportunity arises.
nah77856_ch01.indd 40
17/12/14 6:03 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 41
1.11
Capacity Growth Planning: A Long-Term Strategic Problem 41
Consider the case where the demand exhibits an increasing linear trend. Two policies, (a) and (b), are represented in Figure 1–12. In both cases the firm is acquiring
new capacity at equally spaced intervals x, 2x, 3x, . . . , and increasing the capacity by
the same amount at each of these times. However, in case (a) capacity leads demand,
meaning that the firm maintains excess capacity at all times; whereas in case (b), the
capacity lags the demand, meaning that the existing capacity is fully utilized at all
times. Following policy (a) or (b) results in the time path of excess capacity (or capacity shortfall, if appropriate) given in Figure 1–13.
FIGURE 1–12
xD
Demand
Slope = D per year
x
Number of units
Number of units
Capacity planning
strategies
x
x
Time
Time
(a) Capacity leads demand
(b) Capacity lags demand
FIGURE 1–13
Number of units
Number of units
Time path of excess or
deficient capacity
xD
Time
Time
(a) Capacity leads demand
nah77856_ch01.indd 41
xD
(b) Capacity lags demand
17/12/14 6:03 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 42
nah77856
42 Chapter One Strategy and Competition
Consider the following specific model that appears in Manne (1967). Define
D Annual increase in demand.
x Time interval between introduction of successive plants.
r Annual discount rate, compounded continuously.
f (y) Cost of opening a plant of capacity y.
From Figure 1–12a (which is the strategy assumed in the model), we see that if the
time interval for plant replacement is x, it must be true that the plant size at each
replacement is xD. Furthermore, the present value of a cost of $1 incurred t years into
the future is given by ert. (A discussion of discounting and the time value of money
appears in Appendix 1–A.)
Define C(x) as the sum of discounted costs for an infinite horizon given a plant
opening at time zero. It follows that
C(x) f (x D) erxf (x D) e2rxf (x D) f (x D)[1 erx (erx )2 (erx )3 ]
f(xD)
.
1 e rx
Experience has shown that a representation of f(y) that explains the economies of
scale for plants in a variety of industries is
f( y) kya,
where k is a constant of proportionality. The exponent a measures the ratio of the
incremental to the average costs of a unit of plant capacity. A value of 0.6 seems to
be common (known as the six-tenths rule). As long as a 1, there are economies of
scale in plant construction, since a doubling of the plant size will result in less than a
doubling of the construction costs. To see this, consider the ratio
f(2y) k(2y)a
2a.
f(y)
k(y)a
Substituting a 0.6, we obtain 2a 1.516. This means that if a 0.6 is accurate,
the plant capacity can be doubled by increasing the dollar investment by about 52 percent. Henceforth, we assume that 0 a 1 so that there are economies of scale in the
plant sizing.
Given a specific form for f (y), we can solve for the optimal timing of plant additions
and hence the optimal sizing of new plants. If f ( y) ky a, then
C(x) k(xD)a
.
1 e rx
Consider the logarithm of C(x):
log[C(x)] log[k(xD)a] log[1 erx]
log(k) a log(xD) log[1 erx].
It can be shown that the cost function C(x) has a unique minimum with respect to x
and furthermore that the value of x for which the derivative of log[C(x)] is zero is the
nah77856_ch01.indd 42
17/12/14 6:03 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 43
1.11
FIGURE 1–14
Capacity Growth Planning: A Long-Term Strategic Problem 43
1
The function
u(e u 1)
f(u)
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
u
value of x that minimizes C(x). It is easy to show1 that the optimal solution satisfies
rx
a.
e 1
rx
The function f(u) u(eu 1) appears in Figure 1–14, where u rx. By locating the
value of a on the ordinate axis, one can find the optimal value of u on the abscissa axis.
Example 1.3
A chemicals firm is planning for an increase of production capacity. The firm has estimated that
the cost of adding new capacity obeys the law
f(y) .0107y .62,
where cost is measured in millions of dollars and capacity is measured in tons per year. For
example, substituting y 20,000 tons gives f(y) $4.97 million plant cost. Furthermore, suppose that the demand is growing at a constant rate of 5,000 tons per year and future costs are
discounted using a 16 percent interest rate. From Figure 1–14 we see that, if a .62, the value
of u is approximately 0.9. Solving for x, we obtain the optimal timing of new plant openings:
x ur .9.16 5.625 years.
The optimal value of the plant capacity should be x D (5.625)(5,000) 28,125 tons.
Substituting y 28,125 into the equation for f(y) gives the cost of each plant at the optimal
solution as $6.135 million.
Much of the research into the capacity expansion problem consists of extensions
of models of this type. This particular model could be helpful in some circumstances
1d
log[C(x)]
aD
( e rx)( r)
a
r
rx
0,
dx
xD
1 e rx
x
e 1
which gives
nah77856_ch01.indd 43
rx
a.
erx 1
17/12/14 6:03 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 44
nah77856
44 Chapter One Strategy and Competition
but ignores a number of fundamental features one would expect to find in the real
world:
1. Finite plant lifetime. The assumption of the model is that, once constructed, a plant
has an infinite lifetime. However, companies close plants for a variety of reasons: Equipment becomes obsolete or unreliable and cannot be replaced easily. Labor costs or
requirements may dictate moving either to less expensive locations domestically or to
overseas locations. Major changes in the process technology may not be easily adaptable
to existing facilities.
2. Demand patterns. We have assumed that demand grows at a constant rate per
year. Models have been proposed to account for more complex growth patterns of demand. In truth, demand uncertainty is a key factor. In many industries, foreign competition has made significant inroads into established markets, thus forcing rethinking of
earlier strategies.
3. Technological developments. The model assumes that the capacity of all new
plants constructed remains constant and that the cost of building a plant of given size
remains constant as well. This is obviously unreasonable. Major changes in process
technology occur on a regular basis, changing both the maximum size of new plants
and the costs associated with a fixed plant size.
4. Government regulation. Environmental and safety restrictions may limit choices
of plant location and scale.
5. Overhead costs. Most capacity expansion and location models do not explicitly
account for the costs of overhead. During the energy crunch in the late 1970s, costs of
energy overhead soared, wreaking havoc with plant overhead budgets.
6. Tax incentives. The financial implications of the sizing and location of new facilities must be considered in the context of tax planning. Tax incentives are offered by
local or state municipalities to major corporations considering sites for the construction of new facilities.
An interesting question is whether models of this type really capture the way that
companies have made capacity expansion decisions in the past. There is some preliminary evidence that they do not. Lieberman (1987) attempted to assess the factors that
motivated firms to construct new chemical plants during the period 1957 to 1982. He
found that the size of new plants increased by about 8 percent per year independent of
market conditions. In periods of high demand, firms constructed more plants. This preliminary study indicates that the rational thinking that leads to models such as the one
developed in this section does not accurately reflect the way that companies make
plant-sizing decisions. His results suggest that firms build the largest plants possible
with the existing technology. (Given sufficient economies of scale, however, this policy may theoretically be optimal.)
Issues in Plant Location
This section has been concerned with capacity expansion decisions, specifically, determining the amount and timing of new capacity additions. A related issue is the
location of the new facility. Deciding where to locate a plant is a complex problem.
Many factors must be carefully considered by management before making the final
choice.
The following information about the plant itself is relevant to the location decision:
1. Size of the plant. This includes the required acreage, the number of square feet of
space needed for the building structure, and constraints that might arise as a result
of special needs.
nah77856_ch01.indd 44
17/12/14 6:03 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
Page 45
1.11
Capacity Growth Planning: A Long-Term Strategic Problem 45
2. Product lines to be produced.
3. Process technology to be used.
4. Labor force requirements. These include both the number of workers required and
the specification of the particular skills needed.
5. Transportation needs. Depending on the nature of the product produced and the
requirements for raw materials, the plant may have to be located near major interstate highways or rail lines.
6. Utilities requirements. These include special needs for power, water, sewage, or
fossil fuels such as natural gas. Plants that have unusual power needs should be
located in areas where energy is less expensive or near sources of hydroelectric
power.
7. Environmental issues. Because of government regulations, there will be few
allowable locations if the plant produces significant waste products.
8. Interaction with other plants. If the plant is a satellite of existing facilities, it is
likely that management would want to locate the new plant near the others.
9. International considerations. Whether to locate a new facility domestically or
overseas is a very sensitive issue. Although labor costs may be lower in some
locations, such as the Far East, tariffs, import quotas, inventory pipeline costs, and
market responsiveness also must be considered.
10. Tax treatment. Tax consideration is an important variable in the location decision.
Favorable tax treatment is given by some countries, such as Ireland, to encourage
new industry. There are also significant differences in state tax laws designed to
attract domestic manufacturers.
Mathematical models are useful for assisting with many operational decisions. However, they generally are of only limited value for determining a suitable location for a new
plant. Because so many factors and constraints enter into the decision process, such decisions are generally made based on the inputs of one or more of the company’s divisions,
and the decision process can span several years. (Mathematical techniques for making location decisions will be explored in Chapter 11, on facilities layout and location.)
Schmenner (1982) has examined the decision process at a number of the Fortune 500
companies. His results showed that in most firms, the decision of where to locate a new
facility was made either by the corporate staff or by the CEO, even though the request for
new facilities might have originated at the division level. The degree of decision-making
autonomy enjoyed at the division level depended on the firm’s management style.
Based on a sample survey, Schmenner reported that the major factors that influenced
new facilities location decisions were the following:
1. Labor costs. This was a primary concern for industries such as apparel, leather,
furniture, and consumer electronics. It was less of a concern for capital-intensive
industries.
2. Unionization. A motivating factor for a firm considering expanding an existing
facility, as opposed to one considering building a new facility, is the potential for
eliminating union influence in the new facility. A fresh labor force may be more
difficult to organize.
3. Proximity to markets. When transportation costs account for a major portion of the
cost of goods sold, locating new plants near existing markets is essential.
4. Proximity to supplies and resources. The decision about where to locate plants in
certain industries is based on the location of resources. For example, firms producing wood or paper products must be located near forests and firms producing
processed food, near farms.
nah77856_ch01.indd 45
17/12/14 6:03 PM
nah77856_ch01.qxd
12/28/07
3:46 PM
nah77856
Page 46
46 Chapter One Strategy and Competition
5. Proximity to other facilities. Many companies tend to place manufacturing divisions
and corporate facilities in the same geographic area. For example, IBM originated
in Westchester County in New York, and located many of its divisions in that state.
By locating key personnel near each other, the firm has been able to realize
economies of scope.
6. Quality of life in the region. When other issues do not dictate the choice of a location, choosing a site that will be attractive to employees may help in recruiting key
personnel. This is especially true in high-tech industries that must compete for
workers with particular skills.
Problems for Section 1.11
38. A start-up company, Macrotech, plans to produce a device to translate Morse code
to a written message on a home computer and to send written messages in Morse
code over the airwaves. The device is primarily of interest to ham radio enthusiasts. The president, Ron Lodel, estimates that it would require a $30,000 initial
investment. Each unit costs him $20 to produce and each sells for $85.
a. How many units must be sold in order for the firm to recover its initial
investment?
b. What is the total revenue at the break-even volume?
c. If the price were increased to $100 each, find the break-even volume.
39. For Problem 35, suppose that sales are expected to be 100 units in the first year and
increase at a rate of 40 percent per year. How many years will it take to recoup the
$30,000 initial investment? Assume that each unit sells for $85.
40. A domestic producer of baby carriages, Pramble, buys the wheels from a company
in the north of England. Currently the wheels cost $4 each, but for a number of
reasons the price will double. In order to produce the wheels themselves, Pramble
would have to add to existing facilities at a cost of $800,000. It estimates that its unit
cost of production would be $3.50. At the current time, the company sells 10,000
carriages annually. (Assume that there are four wheels per carriage.)
a. At the current sales rate, how long would it take to pay back the investment
required for the expansion?
b. If sales are expected to increase at a rate of 15 percent per year, how long will
it take to pay back the expansion?
41. Based on past experience, a chemicals firm estimates that the cost of new capacity
additions obeys the law
f (y) .0205y.58
where y is measured in tons per year and f (y) in millions of dollars. Demand is
growing at the rate of 3,000 tons per year, and the accounting department recommends a rate of 12 percent per year for discounting future costs.
a. Determine the optimal timing of plant additions and the optimal size of each
addition.
b. What is the cost of each addition?
c. What is the present value of the cost of the next four additions? Assume an
addition has just been made for the purposes of your calculation. (Refer to
Appendix 1–A for a discussion of cost discounting.)
nah77856_ch01.indd 46
17/12/14 6:03 PM
nah77856_ch01.qxd
1/17/08
3:01 PM
Page 47
1.12
Summary 47
42. A major oil company is considering the optimal timing for the construction of new
refineries. From past experience, each doubling of the size of a refinery at a single
location results in an increase in the construction costs of about 68 percent. Furthermore, a plant size of 10,000 barrels per day costs $6 million. Assume that the
demand for the oil is increasing at a constant rate of two million barrels yearly and
the discount rate for future costs is 15 percent.
a. Find the values of k and a assuming a relationship of the form f(y) kya.
Assume that y is in units of barrels per day.
b. Determine the optimal timing of plant additions and the optimal size of each plant.
c. Suppose that the largest single refinery that can be built with current technology is 15,000 barrels per day. Determine the optimal timing of plant additions
and the optimal size of each plant in this case. (Assume 365 days per year for
your calculations.)
1.12 Summary
nah77856_ch01.indd 47
This chapter discussed the importance of operations strategy and its relationship to
the overall business strategy of the firm. Operations continues to grow in importance in the
firm. While a larger portion of direct manufacturing continues to move off-shore, the importance of the manufacturing function should not be underestimated. The success of the
operations strategy can be measured along several dimensions. These include the obvious
measures of cost and product characteristics, but also include quality, delivery speed,
delivery reliability, and flexibility.
The classical view of manufacturing strategy, due primarily to Wickham Skinner,
considers the following four dimensions of strategy: time horizon, focus, evaluation,
and consistency. Different types of decisions relate to different time frames. A plant should
be designed with a specific focus in mind, whether it be to minimize unit cost or to maximize product quality. Several evaluation criteria may be applied to analyze the effectiveness
of a strategy.
We hear more and more frequently that we are part of a global community. When buying products today, we are less concerned with the country of origin than with the characteristics of the product. How many consumers of cell phones are even aware that Nokia is
headquartered in Finland, Ericsson in Sweden, and Motorola in the United States? An interesting question explored by Michael Porter is: Why do some industries seem to thrive in
some countries? While the answer is complex, Porter suggests that the following four factors are most important: factor conditions; demand conditions; related and supporting
industries; and firm strategy, structure, and rivalry.
Changing the way that one does things can be difficult. Even more difficult is changing the way that a company does things. For that reason, business process engineering
(BPR) is a painful process, even when it works. The most dramatic successes of BPR have
come in service functions, but the concept can be applied to any environment. It is the
process of rethinking how and why things are done in a certain way. Intelligently done,
BPR can lead to dramatic improvements. However, it can also be a time-consuming and
costly process.
Just-in-time (JIT) is a philosophy that grew from the kanban system developed by
Toyota. At the heart of the approach is the elimination of waste. Systems are put in place to
reduce material flows to small batches to avoid large buildups of work-in-process
inventories. While JIT developed on the factory floor, it is a concept that has been applied to
the purchasing function as well. Successful application of JIT purchasing requires the
development of long-term relationships, and usually requires close proximity to suppliers.
The mechanics of JIT are discussed in more detail in Chapter 8.
17/12/14 6:03 PM
nah77856_ch01.qxd
1/25/08
3:59 PM
Page 48
48 Chapter One Strategy and Competition
Being able to get to the market quickly with products that people want in the volumes that
the marketplace requires is crucial if one wants to be a market leader. Time-based competition means that the time from product conception to its appearance in the marketplace
must be reduced. To do so, one performs as many tasks concurrently as possible. In many instances, time to market is less important than time to volume. Being the first to the market
may not mean much if one cannot meet product demand.
The dramatic successes of the Japanese during the 1970s and 1980s were to a large
extent due to the outstanding quality of their manufactured products. Two Americans,
Deming and Juran, visited Japan in the early 1950s and played an important role in making
the Japanese aware of the importance of producing quality products. The quality movement
in the United States has resulted in a much greater awareness of the importance of quality,
recognition for outstanding achievement in this arena via the Malcolm Baldrige Award, and
initiation of important programs such as quality circles and the six-sigma program at
Motorola. (Both the statistical and the organizational issues concerning quality are discussed in detail in Chapter 12.)
It is important to understand both product and process life cycles. Both go through the
four cycles of start-up, rapid growth, maturation, and stabilization or decline. It is also important to understand which types of processes are appropriate for which types of products
and industries. To this end, Hayes and Wheelwright have developed the concept of the
product–process matrix.
Learning and experience curves are useful in modeling the decline in labor hours or
the decline in product costs as experience is gained in the production of an item or family
of items. These curves have been shown to obey an exponential law, and can be useful predictors of the cost or time required for production. (Moore’s Law, due to Gordon Moore, a
founder of Intel, predicted the doubling of chip performance every 18 months. This is
an example of an experience curve, and the prediction has continued to be accurate to the
present day.)
We discussed two methods for assisting with capacity expansion decisions. Breakeven curves provide a means of determining the sales volume necessary to justify
investing in new or existing facilities. A simple model for a dynamic expansion policy is
presented that gives the optimal timing and sizing of new facilities assuming constant
demand growth and discounting of future costs. We also discussed issues that arise in
trying to decide where to locate new facilities. This problem is very complex in that there
are many factors that relate to the decision of where to locate production, design, and
management facilities.
Additional
Problems for
Chapter 1
nah77856_ch01.indd 48
43. What is a production and operations strategy? Discuss the elements in common with
marketing and financial strategies and the elements that are different.
44. What is the difference between the product life cycle and the process life cycle?
In what way are these concepts related?
45. Suppose that the Mendenhall Corporation, a producer of women’s handbags, has
determined that a 73 percent experience curve accurately describes the evolution of its
production costs for a new line. If the first unit costs $100 to produce, what should the
10,000th unit cost based on the experience curve?
46. Delon’s Department Store sells several of its own brands of clothes and several
well-known designer brands as well. Delon’s is considering building a plant in
17/12/14 6:03 PM
nah77856_ch01.qxd
1/25/08
3:59 PM
Page 49
Additional Problems for Chapter 1 49
Malaysia to produce silk ties. The plant will cost the firm $5.5 million. The
plant will be able to produce the ties for $1.20 each. On the other hand, Delon’s
can subcontract to have the ties produced and pay $3.00 each. How many ties
will Delon’s have to sell worldwide to break even on its investment in the new
plant?
47. A Japanese steel manufacturer is considering expanding operations. From experience, it
estimates that new capacity additions obey the law
f ( y) .00345y.51,
where the cost f ( y) is measured in millions of dollars and y is measured in tons of
steel produced. If the demand for steel is assumed to grow at the constant rate of
8,000 tons per year and future costs are discounted using a 10 percent discount rate,
what is the optimal number of years between new plant openings?
The following problems are designed to be solved by spreadsheet.
48. Consider the following break-even problem: the cost of producing Q units, c(Q), is
described by the curve
c(Q) 48Q[1 exp(.08Q)],
where Q is in hundreds of units of items produced and c(Q) is in thousands of
dollars.
a. Graph the function c(Q). What is its shape? What economic phenomenon gives rise
to a cumulative cost curve of this shape?
b. At what production level does the cumulative production cost equal
$1,000,000?
c. Suppose that these units can be purchased from an outside supplier at a cost of $800
each, but the firm must invest $850,000 to build a facility that would be able to
produce these units at a cost c(Q). At what cumulative volume of production does
it make sense to invest in the facility?
49. Maintenance costs for a new facility are expected to be $112,000 for the first year of
operation. It is anticipated that these costs will increase at a rate of 8 percent per year.
Assuming a rate of return of 10 percent, what is the present value of the stream of
maintenance costs over the next 30 years?
50. Suppose the supplier of keyboards described in Example 1.2 is willing to offer the
following incremental quantity discount schedule:
Cost per
Keyboard
$50
$45
$40
Order Quantity
Q 100,000
100,000 Q 500,000
500,000 Q
Determine the cost to the firm for order quantities in increments of 20,000
for Q 200,000 to Q 1,000,000, and compare that to the cost to the firm of
producing internally for these same values of Q. What is the break-even order
quantity?
nah77856_ch01.indd 49
17/12/14 6:03 PM
nah77856_ch01.qxd
1/17/08
3:01 PM
Page 50
50 Chapter One Strategy and Competition
Appendix 1–A
PRESENT WORTH CALCULATIONS
Including the time value of money in the decision process is common when considering
alternative investment strategies. The idea is that a dollar received today has greater value
than one received a year from now. For example, if a dollar were placed in a simple passbook account paying 5 percent, it would be worth $1.05 in one year. More generally, if it
were invested at a rate of return of r (expressed as a decimal), it would be worth 1 r in a
year, (1 r)2 in two years, and so on.
In the same way, a cost of $1 incurred in a year has a present value of less than $1 today.
For example, at 5 percent interest, how much would one need to place in an account today
so that the total principal plus interest would equal $1 in a year? The answer is 1(1.05) 0.9524. Similarly, the present value of a $1 cost incurred in two years at 5 percent is
1(1.05)2 0.9070. In general, the present value of a cost of $1 incurred in t years assuming a rate of return r is (1 r)t.
These calculations assume that there is no compounding. Compounding means that one
earns interest on the interest, so to speak. For example, 5 percent compounded semiannually means that one earns 2.5 percent on $1 after six months and 2.5 percent on the original
$1 plus interest earned in the first six months. Hence, the total return is
(1.025)(1.025) $1.050625
after one year, or slightly more than 5 percent. If the interest were compounded quarterly,
the dollar would be worth
(1 .054)4 1.0509453
at the end of the year. The logical extension of this idea is continuous compounding. One
dollar invested at 5 percent compounded continuously would be worth
lim (1 .05n)n e.05
n→
1.05127
at the end of a year. The number e 2.7172818 . . . is defined as
e lim (1 1n)n.
n→
Notice that continuous compounding only increases the effective simple interest rate
from 5 percent to 5.127 percent.
More generally, C invested at a rate of r for t years compounded continuously is worth
Cert at the end of t years.
Reversing the argument, the present value of a cost of C incurred in t years assuming
continuous compounding at a discount rate r is Cert. A stream of costs C1, C2, . . . , Cn
incurred at times t1, t2, . . . , tn has present value
a Ci e
n
rti
.
i1
A comprehensive treatment of discounting and its relationship to the capacity expansion
problem can be found in Freidenfelds (1981).
nah77856_ch01.indd 50
17/12/14 6:03 PM
Bibliography 51
Bibliography
Abernathy, W. J., and P. L. Townsend. “Technology, Productivity, and Process Change.” Technological Forecasting and
Social Change 7 (1975), pp. 379–96.
Abernathy, W. J., and K. Wayne. “Limits of the Learning
Curve.” Harvard Business Review 52 (September–October
1974), pp. 109–19.
Becker, G. S. Quoted in Business Week, January 27, 1986,
p. 12.
Bell, D. The Coming of the Post-Industrial Society: A
­Venture in Social Forecasting. New York: Basic Books,
1987.
Blackburn, J. D. Time-Based Competition: The Next
­Battleground in American Manufacturing. New York:
McGraw-Hill/Irwin, 1991.
Business Week. October 22, 1990, pp. 94–95.
Business Week. “The Quality Imperative.” October 25, 1991.
Cohen, S. S., and J. Zysman. Manufacturing Matters: The
Myth of the Post-Industrial Economy. New York: Basic
Books, 1987.
Davis, D. “Beating the Clock.” Electronic Business,
May 1989.
Devinney, T. M. “Entry and Learning.” Management
­Science 33 (1987), pp. 102–12.
Dhalla, N. K., and S. Yuspeh. “Forget about the P
­ roduct
Life Cycle Concept.’’ Harvard Business Review 54
(­January–February 1976), pp. 102–12.
The Economist. “Reshoring Manufacturing: Coming Home.”
January 2013. Retrieved from http://www.economist.com/
news/special-report/21569570-growing-number-americancompanies-are-moving-their-manufacturing-back-united
Fallon, M. San Jose Mercury News, September 30, 1985,
p. 11D.
Faux, J. “Manufacturing Key to America’s Future.”
­Presentation to the Industrial Union Council Legislative
Conference, Economic Policy Institute, 7 pages,
February 4, 2003.
Freeland, J. R. “A Survey of Just-in-Time Purchasing
­Practices in the United States.” Production and Inventory
Management Journal, Second Quarter 1991, pp. 43–50
Freidenfelds, J. Capacity Expansion: Analysis of S­ imple
Models with Applications. New York: Elsevier North
­Holland, 1981.
Goldhar, J. P., and M. Jelinek. “Plan for Economies of
Scope.” Harvard Business Review 61 (1983), pp. 141–48.
Hagenbaugh, B. “U.S. Manufacturing Jobs Fading Away
Fast.” USA Today, December 12, 2002.
Hammer, M. S., and J. Champy. Reengineering the
­Corporation: A Manifesto for Business Resolution.
New York: Harper Business, 1993.
Hammer, M. S., and S. A. Stanton. The Reengineering
­Revolution. New York: Harper Business, 1995.
nah77856_ch01.indd 51
Hayes, R. H., and S. Wheelwright. “Link Manufacturing
Profess and Product Life Cycles.” Harvard Business Review
57 (January–February 1979), pp. 133–40.
Hayes, R. H., and S. Wheelwright. Restoring Our
­Competitive Edge: Competing through Manufacturing.
New York: John Wiley & Sons, 1984.
Hutzel, T. and D. Lippert. Reshoring 101: Rebuilding U.S.
Manufacturing through Right Sizing and Right Shoring.
­Retrieved from http://www.mainstreammanagement.com/
pdf/rebuilding-us-manufacturing.pdf, 2012.
Jones, P., and L. Kahaner. Say It and Live It. The Fifty
­Corporate Mission Statements That Hit the Mark.
New York: Currency Doubleday, 1995.
Koerner, B. I. “Made in America: Small Businesses Buck
the Offshoring Trend.” Wired, March 2011, pp. 104–111.
Krugman, P. Peddling Prosperity: Economic Sense and
Nonsense in the Age of Diminished Expectations. New York:
W. W. Norton and Company, 1994.
Lieberman, M. “Scale Economies, Factory Focus, and Optimal Plant Size.” Paper present at Stanford Graduate School
of Business, July 22, 1987.
Manne, A. S., ed. Investments for Capacity Expansion: Size,
Location, and Time Phasing. Cambridge, MA: MIT Press,
1967.
Noyce, R. N. “Microelectronics.” Scientific American,
­September 1977.
Panzer, J. C., and R. O. Willing. “Economies of Scope.”
American Economic Review, 1981, pp. 268–72.
Pisano, G.P. and W.C. Shih. Producing Prosperity: Why
America Needs a Manufacturing Renaissance. Harvard
Business Review Press, Boston, Massachusetts, 2012.
Port, O. “Customers Move into the Driver’s Seat.” Business
Week. October 4, 1999, pp. 103–06.
Porter, M. E. and J. W. Rivkin. “Choosing the United
States.” Harvard Business Review, pp. 80–93, March 2012.
Shen, Y. “Selection Incentives in a Performance-Based
­Contracting System.” Health Services Research. 38 (2003),
pp. 535–552.
Skinner, W. Manufacturing in the Corporate Strategy.
New York: John Wiley & Sons, 1978.
Stasey, R., and C. J. McNair. Crossroads: A JIT Success
Story. New York: McGraw-Hill/Irwin, 1990.
Vandermerwe, S. and J. Rada. “Servitization of Business:
Adding Value by Adding Services.” European Management
Journal 6 (1988), pp. 314–324.
Wills, F. “Bully for Taurus.” Business Month, February
1990, p. 13.
Wise, R. and P. Baumgartner. “Go Downstream: The New
Profit Imperative in Manufacturing.” Harvard Business
­Review 77 (1999), pp. 133–141.
Womack, J. P., D. T. Jones, and D. Roos. The Machine That
Changed the World. New York: Harper Perennial, 1990.
17/12/14 6:03 PM
Chapter Two
Forecasting
“It’s hard to make predictions, especially about the future.”
—Neils Bohr
Chapter Overview
Purpose
To present and illustrate the most important methods for forecasting demand in the
context of operations planning.
Key Points
1. Characteristics of forecasts.
• They are almost always going to be wrong.
• A good forecast also gives some measure of error.
• Forecasting aggregate units is generally easier than forecasting individual units.
• Forecasts made further out into the future are less accurate.
• A forecasting technique should not be used to the exclusion of known
­information.
2. Subjective forecasting. Refers to methods that measure either individual or group
opinion. The better known subjective forecasting methods include:
• Sales force composites.
• Customer surveys.
• Jury of executive opinion.
• The Delphi method.
3. Objective forecasting methods (time series methods and regression). Using objective forecasting methods, one makes forecasts based on past history. Time series
forecasting uses only the past history of the series to be forecasted, while regression
models often incorporate the past history of other series. In time series forecasting,
the goal is to find predictable and repeatable patterns in past data. Based on the
identified pattern, different methods are appropriate. Time series methods have
the advantage of easily being incorporated into a computer program for automatic
forecasting and updating. Repeatable patterns that we consider include increasing
or decreasing linear trend, curvilinear trend (including exponential growth), and
seasonal fluctuations. When using regression, one constructs a causal model that
predicts one phenomenon (the dependent variable) based on the evolution of one
or more other phenomenon (the independent variables). An example would be
predicting the start or end of a recession based on housing starts (housing starts are
considered to be a leading economic indicator of the health of the economy).
52
nah77856_ch02.indd 52
17/12/14 6:47 PM
Chapter Two Forecasting 53
4. Evaluation of forecasting methods. The forecast error in any period, et, is the
difference between the forecast for period t and the actual value of the series
realized for period t (et 5 Ft 2 Dt). Three common measures of forecast error are
MAD (average of the absolute errors over n periods), MSE (the average of the sum
of the squared errors over n periods), and MAPE (the average of the percentage
errors over n periods).
5. Methods for forecasting stationary time series. We consider two forecasting methods when the underlying pattern of the series is stationary over time:
moving averages and exponential smoothing. A moving average is simply the
arithmetic average of the N most recent observations. Exponential smoothing
forecasts rely on a weighted average of the most recent observation and the
previous forecast. The weight applied to the most recent observation is a, where
0 , a , 1, and the weight applied to the last forecast is 1 2 a. Both methods
are commonly used in practice, but the exponential smoothing method is favored
in inventory control applications—especially in large systems—because it requires
much less data storage than does moving averages.
6. Methods for forecasting series with trend. When there is an upward or downward
linear trend in the data, two common forecasting methods are linear regression
and double exponential smoothing via Holt’s method. Linear regression is used
to fit a straight line to past data based on the method of least squares, and Holt’s
method uses separate exponential smoothing equations to forecast the intercept
and the slope of the series each period.
7. Methods for forecasting seasonal series. A seasonal time series is one that
has a regular repeating pattern over the same time frame. Typically, the
time frame would be a year, and the periods would be weeks or months. The
simplest approach for forecasting seasonal series is based on multiplicative
seasonal factors. A multiplicative seasonal factor is a number that indicates
the relative value of the series in any period compared to the average value
over a year. Suppose a season consists of 12 months. A seasonal factor of
1.25 for a given month means that the demand in that month is 25 percent
higher than the mean monthly demand. Winter’s method is a more complex
method based on triple exponential smoothing. Three distinct smoothing
equations are used to forecast the intercept, the slope, and the seasonal
­factors each period.
8. Box-Jenkins models. George Box and Gwilym Jenkins developed forecasting
methods based largely on a statistical analysis of the autoregressive function
of a time series. Autoregression seeks to discover repeating patterns in data by
considering the correlation of observations of the series with other observations
separated by fixed number of periods. These models have proven to be very powerful for forecasting some economic time series, but they require large data sets
(at least 72 observations) and a knowledgeable user. We provide a brief review of
these powerful methods.
9. Other considerations. In addition to Box-Jenkins methods, when large data sets
are available, filtering methods borrowed from electrical engineering can often
provide excellent forecasts for economic time series. Two of the better known
filters are Kalman Filters and Wiener Filters. Neither of these methods are
amenable to automatic forecasting. Monte Carlo simulation is another technique
that can be useful for building a forecasting model. Finally, we discuss forecasting
demand in the context of a lost sales inventory system.
nah77856_ch02.indd 53
17/12/14 6:47 PM
54 Chapter Two Forecasting
Families plan vacations around their schedules, and for that reason America’s theme
parks tend to be very crowded during holidays and school breaks. On June 18, 2011,
Universal Studios Theme Parks opened its new Harry Potter section at Universal
­Orlando Resort in Orlando, FL, which continues to be a major attraction for ­tourists.
If you were planning on taking your family to a theme park such as Universal ­Studios,
when would be the best time to go to avoid the long lines? One might think that
Thanksgiving and Christmas would be good choices, as most people are home with
families on these holidays. Not so! In fact, these two days (along with New Year’s
Day) are the busiest days of the year. The period between Thanksgiving and Christmas
as well as the periods after New Year’s Day and before Halloween are the slowest
times, and Universal Orlando Resort recommends that visitors come at these times to
avoid longer lines.
As was so eloquently stated by Charles F. Kettering, “My concern is with the ­future
since I plan to spend the rest of my life there.” But the future can never be known,
so we make forecasts. We forecast traffic patterns and plan routes accordingly. We
­forecast which foods will be best in a particular restaurant, and order accordingly. We
choose universities to attend based on forecasting our experiences there and the doors
that a degree from that university will open. We make hundreds of forecasts every
day, some carefully thought out, some made almost unconsciously. Forecasting plays
a central role in all of our lives.
In the same way, forecasting plays a central role in the operations function of a
firm. All business planning is based on forecasts. Sales of existing and new products,
requirements and availabilities of raw materials, changing skills of workers, interest
rates, capacity requirements, and international politics are only a few of the factors
likely to affect the future success of a firm.
The functional areas of the firm that make the most use of forecasting methods are
marketing and production. Marketing is responsible for forecasting sales of both new
and existing product lines. Sales forecasts are the primary driver for the S&OP (sales
and operations planning) function, which will be discussed in detail in Chapter 3.
In some circumstances, the forecasts prepared for marketing purposes may not be
appropriate or sufficient for operations planning. For example, to determine suitable
stocking levels for spare parts, one must know schedules for planned ­replacements
and be able to forecast unplanned replacements. Also it could be that the S&OP
­planning function might be producing forecasts for aggregate units, while forecasts for
­individual SKU’s (stock-keeping units) might be required.
We have seen firms benefit from good forecasting and pay the price for poor
forecasting. During the 1960s, consumer tastes in automobiles slowly shifted from
large, heavy gas guzzlers to smaller, more fuel efficient automobiles. Detroit, slow to
­respond to this change, suffered when the OPEC oil embargo hit in the late 1970s and
tastes shifted more dramatically to smaller cars. Compaq Computer became a market
leader in the early 1980s by properly forecasting consumer demand for a portable
version of the IBM PC, which gained a popularity that far exceeded expectations.
Forecasting played a role in Ford Motors’s early success and later demise. Henry
Ford saw that the consumer wanted a simpler, less expensive car that was easier to
­maintain than most manufacturers were offering in the early 1900s. His Model T
dominated the industry. However, Ford did not see that consumers would tire of the
open Model T design. Ford’s failure to forecast consumer desires for other designs
nearly resulted in the end of a firm that had monopolized the industry only a few
years before.
nah77856_ch02.indd 54
17/12/14 6:47 PM
2.1 The Time Horizon in Forecasting 55
Seeing trends is the first step towards profiting from those trends. As an example,
consider the trend towards greater use of renewable energy. Renewable energy sources
include wind power, sun power, tidal power, geothermal power, etc. If energy can
be generated and stored using renewable methods, this energy can be used to power
electric cars, thus cutting down on gasoline consumption.
Some companies were able to see this trend and take advantage of it. In particular, the use of solar cells has grown dramatically in recent years. While Apple
Corporation has received a great deal of publicity for its fantastic successes in
mobile computing, solar cell installations have actually been growing at a comparable rate. The residential use of solar cells in the United States grew 33 percent
in Q1 2013 compared to Q1 2012. Markets in Asia are expected to grow more
rapidly than in the United States, thus assuring a steady market growth in this segment. Manufacturers that saw this trend developing are now reaping the rewards
of their foresight.
Can all events be accurately forecasted? The answer is clearly no. Consider the
­experiment of tossing a coin. Assuming that it is a fair coin and the act of tossing
does not introduce bias, the best you can say is that the probability of getting heads is
50 percent on any single toss. No one has been able to consistently top the 50 percent
prediction rate for such an experiment over a long period of time. Many real phenomena are accurately described by a type of coin-flipping experiment. Games of chance
played at casinos are random. By tipping the probabilities in its favor, the house is
­always guaranteed to win over the long term. There is evidence that daily prices of
stocks follow a purely random process, much like a coin-flipping experiment. Studies have shown that professional money managers rarely outperform stock portfolios
generated purely at random.
In production and operations management, we are primarily interested in forecasting product demand. Because demand is likely to be random in most circumstances, can forecasting methods provide any value? In most cases, the answer is
yes. Although some portions of the demand process may be unpredictable, other
portions may be predictable. Trends, cycles, and seasonal variation may be present,
all of which give us an advantage over trying to predict the outcome of a coin toss.
In this chapter we consider methods for predicting future values of a series based on
past observations.
2.1 The Time Horizon in Forecasting
We may classify forecasting problems along several dimensions. One is the time
­hori­zon. Figure 2–1 is a schematic showing the three time horizons associated with
­forecasting and typical forecasting problems encountered in operations planning
associated with each. Short-term forecasting is crucial for day-to-day planning.
Short-term forecasts, typically measured in days or weeks, are required for inventory m
­ anagement, production plans that may be derived from a materials requirements planning system (to be discussed in detail in Chapter 8), and resource
requirements planning. Shift scheduling may require forecasts of workers’ availabilities and p­ references.
The intermediate term is measured in weeks or months. Sales patterns for p­ roduct
families, requirements and availabilities of workers, and resource requirements
are ­
typical i­ntermediate-term forecasting problems encountered in o­perations
management.
nah77856_ch02.indd 55
17/12/14 6:47 PM
56 Chapter Two Forecasting
Figure 2–1
Forecast horizons in
operation planning
Short-term sales
Shift schedule
Resource requirements
Da
We
eks
Int
e
ediate
rm
Sho
rt
ys/
Product family sales
Labor needs
Resource requirements
ths
on
M
k s/
e
We
Months/Years
Long
Capacity needs
Long-term sales patterns
Growth trends
Long-term production and manufacturing decisions, discussed in Chapter 1, are
part of the overall firm’s manufacturing strategy. One example is long-term planning
of capacity needs. When demands are expected to increase, the firm must plan for
the construction of new facilities and/or the retrofitting of existing facilities with new
­technologies. Capacity planning decisions may require downsizing in some circumstances. For example, General Motors Corporation historically commanded about
45 percent of the domestic car market. However, in the 1990s that percentage dropped
to 35 percent. As a result, GM was forced to significantly curtail its manufacturing
­operations to remain profitable.
2.2 Characteristics of Forecasts
1. They are usually wrong. As strange as it may sound, this is probably the most
i­gnored and most significant property of almost all forecasting methods. Forecasts,
once determined, are often treated as known information. Resource requirements and
production schedules may require modifications if the forecast of demand proves to be
­inaccurate. The planning system should be sufficiently robust to be able to react to un­
anticipated forecast errors.
2. A good forecast is more than a single number. Given that forecasts are generally
wrong, a good forecast also includes some measure of the anticipated forecast error.
This could be in the form of a range, or an error measure such as the variance of the
distribution of the forecast error.
nah77856_ch02.indd 56
17/12/14 6:47 PM
2.3 Subjective Forecasting Methods 57
3. Aggregate forecasts are more accurate. Recall from statistics that the variance
of the average of a collection of independent identically distributed random variables
is lower than the variance of each of the random variables; that is, the variance of the
sample mean is smaller than the population variance. This same phenomenon is true
in forecasting as well. On a percentage basis, the error made in forecasting sales for
an entire product line is generally less than the error made in forecasting sales for an
­individual item. This phenomenon, known as risk pooling, will be discussed in the
inventory control context in Chapter 5.
4. The longer the forecast horizon, the less accurate the forecast will be. This
­property is quite intuitive. One can predict tomorrow’s value of the Dow Jones
­Industrial Average more accurately than next year’s value.
5. Forecasts should not be used to the exclusion of known information. A ­partic­ular
technique may result in reasonably accurate forecasts in most circumstances. ­However,
there may be information available concerning the future demand that is not presented
in the past history of the series. For example, the company may be planning a special
promotional sale for a particular item so that the demand will probably be higher than
normal. This information must be manually factored into the forecast.
2.3 Subjective Forecasting Methods
We classify forecasting methods as either subjective or objective. A subjective forecasting method is based on human judgment. There are several techniques for soliciting
opinions for forecasting purposes:
1. Sales force composites. In forecasting product demand, a good source of subjective information is the company sales force. The sales force has direct contact with
consumers and is therefore in a good position to see changes in their preferences. To
­develop a sales force composite forecast, members of the sales force submit sales
­estimates of the products they will sell in the coming year. These estimates might be
individual numbers or several numbers, such as pessimistic, most likely, and optimistic estimates. Sales managers would then be responsible for aggregating individual
estimates to arrive at overall forecasts for each geographic region or product group.
Sales force composites may be i­naccurate when compensation of sales personnel is
based on meeting a quota. In that case, there is clearly an incentive for the sales force
to lowball its estimates.
2. Customer surveys. Customer surveys can signal future trends and shifting preference patterns. To be effective, however, surveys and sampling plans must be carefully
designed to guarantee that the resulting data are statistically unbiased and representative
of the customer base. Poorly designed questionnaires or an invalid sampling scheme
may result in the wrong conclusions.
3. Jury of executive opinion. When there is no past history, as with new products, e­ xpert opinion may be the only source of information for preparing forecasts.
The ­approach here is to systematically combine the opinions of experts to derive a
forecast. For new product planning, opinions of personnel in the functional areas
of marketing, ­finance, and operations should be solicited. Combining individual
forecasts may be done in several ways. One is to have the individual responsible
for preparing the f­ orecast interview the executives directly and develop a forecast
from the results of the interviews. Another is to require the executives to meet as a
group and come to a c­ onsensus.
nah77856_ch02.indd 57
17/12/14 6:47 PM
58 Chapter Two Forecasting
4. The Delphi method. The Delphi method, like the jury of executive opinion
method, is based on soliciting the opinions of experts. The difference lies in the
manner in which individual opinions are combined. (The method is named for the
Delphic ­oracle of ancient Greece, who purportedly had the power to predict the future.) The ­Delphi method attempts to eliminate some of the inherent shortcomings
of group ­dynamics, in which the personalities of some group members overshadow
those of other members. The method requires a group of experts to express their
opinions, preferably by individual sample survey. The opinions are then compiled
and a summary of the results is returned to the experts, with special attention to
those opinions that are significantly different from the group averages. The ­experts
are asked if they wish to r­econsider their original opinions in light of the group
response. The process is repeated until (ideally) an overall group consensus is
reached.
As with any particular technique, the Delphi method has advantages and disadvantages. Its primary advantage is that it provides a means of assessing individual opinion
without the usual concerns of personal interactions. On the negative side, the method is
highly sensitive to the care in the formulation of the questionnaire. Because discussions
are intentionally excluded from the process, the experts have no mechanism for resolving ambiguous questions. Furthermore, it is not necessarily true that a group consensus
will ever be reached. An interesting case study of a successful ­application of the ­Delphi
method can be found in Helmer and Rescher (1959).
2.4 Objective Forecasting Methods
Objective forecasting methods are those in which the forecast is derived from
an ­analysis of data. A time series method is one that uses only past values of the
­phenomenon we are predicting. Causal models are ones that use data from sources
other than the series being predicted; that is, there may be other variables with values
that are linked in some way to what is being forecasted. We discuss these first.
Causal Models
Let Y represent the phenomenon we wish to forecast and X1, X2, . . . , Xn be n variables
that we believe to be related to Y. Then a causal model is one in which the forecast for
Y is some function of these variables, say,
Y 5 f(X1, X2, . . . , Xn).
Econometric models are special causal models in which the relationship between
Y and (X1, X2, . . . , Xn) is linear. That is,
Y 5 a0 1 a1 X1 1 a2 X2 1 ? ? ? 1 an Xn
for some constants (al , . . . , an). The method of least squares is most commonly used
to find estimators for the constants. (We discuss the method in Appendix 2–B for the
case of one independent variable.)
Let us consider a simple example of a causal forecasting model. A realtor is trying to
estimate his income for the succeeding year. In the past he has found that his i­ncome is
close to being proportional to the total number of housing sales in his territory. He also
nah77856_ch02.indd 58
17/12/14 6:47 PM
2.4 Objective Forecasting Methods 59
has noticed that there has typically been a close relationship between housing sales and
interest rates for home mortgages. He might construct a model of the form
Yt 5 a0 1 a1Xt21,
where Yt is the number of sales in year t and Xt21 is the interest rate in year t 2 1. Based
on past data he would then determine the least squares estimators for the constants
a0 and al. Suppose that the values of these estimators are currently a0 5 385.7 and
a1 5 21,878. Hence, the estimated relationship between home sales and mortgage
rates is
Yt 5 385.7 2 1,878Xt21,
where Xt21, the previous year’s interest rate, is expressed as a decimal. Then if the current mortgage interest rate is 10 percent, the model would predict that the number of
sales the following year in his territory would be 385.7 2 187.8 5 197.9, or about 198
houses sold.
Causal models of this type are common for predicting economic phenomena
such as the gross national product (GNP) and the gross domestic product (GDP).
Both MIT and the Wharton School of Business at the University of ­Pennsylvania
have developed large-scale econometric models for making these predictions.
Econometric prediction models are typically used by the economics and finance
arms of the firm to forecast values of macroeconomic variables such as interest
rates and currency exchange rates. Time series methods are more commonly used
for operations planning applications.
Time Series Methods
Time series methods are often called naive methods, as they require no information
other than the past values of the variable being predicted. Time series is just a fancy
term for a collection of observations of some economic or physical phenomenon drawn
at discrete points in time, usually equally spaced. The idea is that information can be
inferred from the pattern of past observations and can be used to forecast future values
of the series.
In time series analysis we attempt to isolate the patterns that arise most often. These
include the following:
1. Trend. Trend refers to the tendency of a time series to exhibit a stable pattern
of growth or decline. We distinguish between linear trend (the pattern described by a
straight line) and nonlinear trend (the pattern described by a nonlinear function, such as a
quadratic or exponential curve). When the pattern of trend is not specified, it is generally
understood to be linear.
2. Seasonality. A seasonal pattern is one that repeats at fixed intervals. In time ­series
we generally think of the pattern repeating every year, although daily, weekly, and
monthly seasonal patterns are common as well. Fashion wear, ice cream, and heating
oil exhibit a yearly seasonal pattern. Consumption of electricity exhibits a strong daily
seasonal pattern.
3. Cycles. Cyclic variation is similar to seasonality, except that the length and the
­magnitude of the cycle may vary. One associates cycles with long-term eco­nomic variations (that is, business cycles) that may be present in addition to seasonal fluctuations.
nah77856_ch02.indd 59
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:20 PM
Page 59
Snapshot Application
ADVANCED FORECASTING, INC., SERVES
THE SEMICONDUCTOR INDUSTRY
Advanced Forecasting, Inc. (AFI), is a Cupertino-based
firm that specializes in providing forecasts for semiconductor sales and sales of related industries, such as semiconductor equipment and suppliers. The firm has had a
history of accurately predicting the turning points in the
sales patterns of semiconductors for more than a
decade. Forecasts are determined from quantitative
models (such as the ones discussed in this chapter).
Although the actual models used are proprietary, forecasts are based on basic economic factors related to the
semiconductor industry. According to the firm’s founder
Dr. Moshe Handelsman, the problem with most forecaster’s predictions is that they are based on subjective
opinions and qualitative data. AFI uses a mathematical
model to derive its forecasts, which are not second
guessed. While the firm is only a small player in the
semiconductor forecasting arena, their success has been
dramatic. They have consistently been able to predict
major shifts in the market for semiconductors, which is a
fundamental need for management. According to JeanPhilippe Dauvin, vice president and chief economist
for SGS-Thomson Microelectronics: “Our top management pays more attention to Advanced Forecasting’s
predictions than to any other industry source.” Accurate
forecasts allow management to deal with important
strategic issues such as when production capacity should
be expanded, what personnel needs will be, and what
the demands will be on marketing and sales. The success
of AFI demonstrates that quantitative-based forecasting
can provide consistently accurate forecasts and, over
the long term, are far more reliable than subjective
methods.
Source: Advanced Forecasting, Inc., Website, http://www.
adv-forecast.com/afi/.
4. Randomness. A pure random series is one in which there is no recognizable
pattern to the data. One can generate patterns purely at random that often appear to
have structure. An example of this is the methodology of stock market chartists who
impose forms on random patterns of stock market price data. On the other side of
the coin, data that appear to be random could have a very definite structure. Truly
random data that fluctuate around a fixed mean form what is called a horizontal
pattern.
Examples of time series exhibiting some of these patterns are given in Figure 2–2.
Problems for Sections 2.1–2.4
1. Name the four components of time series (i.e., the four distinct patterns exhibited
by time series).
2. What distinguishes seasonality from cycles in time series analysis?
3. What is the appropriate type of forecasting method to use in each of the following
scenarios:
a. Holiday Inn, Inc., is attempting to predict the demand next year for motel rooms,
based on a history of demand observations.
b. Standard Brands has developed a new type of outdoor paint. The company wishes
to forecast sales based on new housing starts.
c. IBM is trying to ascertain the cost of a stock-out of a critical tape drive component. They do so by sample survey of managers at various national spare parts
centers. The surveys are sent back to the managers for a reassessment, and the
process is repeated until a consensus is reached.
nah77856_ch02.indd 60
17/12/14 6:47 PM
nah77856_ch02.qxd
60
12/31/07
4:20 PM
Page 60
Chapter Two Forecasting
2.4 Objective Forecasting Methods 61
FIGURE 2–2
Increasing
linear trend
Purely random—
No recognizable
pattern
Demand
Demand
Time series patterns
Curvilinear
trend (quadratic,
exponential)
Time
Demand
Demand
Time
Time
Seasonal
pattern plus
linear growth
Time
4. Discuss the role of forecasting for the following functions of the firm:
a. Marketing
b. Accounting
c. Finance
d. Production
5. Distinguish between the following types of forecasts:
a. Aggregate versus single item.
b. Short-term versus long-term.
c. Causal versus naive.
6. What is the advantage of the Delphi method over the jury of executive opinion
method? What do these methods have in common?
7. Consider the problem of choosing an appropriate college to attend when you were
a high school senior. What forecasting concerns did you have when you made that
decision? In particular, list the short-term, intermediate-term, and long-term forecasts you might have considered in making your final decision. What objective
sources of data might you have used to provide you with better forecasts in each
case?
8. Discuss the following quotation from an inventory control manager: “It’s not my
fault we ran out of those parts. The demand forecast was wrong.”
9. Discuss the following statement: “Economists are predicting that interest rates will
continue to be under 10 percent for at least 15 years.”
nah77856_ch02.indd 61
17/12/14 6:47 PM
nah77856_ch02.qxd
nah77856_ch02.qxd
12/31/07
12/31/07
4:20 PM
4:20 PM
Page 61
Page 61
62 Chapter Two Forecasting
2.6 Evaluating Forecasts 61
2.6 Evaluating Forecasts 61
2.5
NOTATION CONVENTIONS
2.5 Notation
Conventions
2.5 NOTATION CONVENTIONS
The following discussion
discussiondeals
dealswith
withtime
timeseries
seriesmethods.
methods.
Define
. ,t, D
..
Define
D1D
,D
. .t,.. as
1,2,D.2., .. ,. D
The
following
discussion
withduring
time
series
D2, will
. . . through, assume
Dt, . . .
as
observed
values
ofdeals
demand
periods
t, . will
.D. 1., assume
We
thethe
observed
values
of demand
during
periods
1, 2,methods.
. .1,. , 2,
t, . .Define
. ., We
as
values
of
demand
during
periods
2,predict.
. .to. ,predict.
t, .­F.urthermore,
. . Furthermore,
We will assume
t �the
1} time
is theseries
time
series
we would
we
throughout
, t ${D
1}t, is
we would
like1,tolike
we will
outthe
thatobserved
{Dtthat
,t�
1}forecasting
is the time
series
would
toobserved
predict.
Furthermore,
we
throughout
{D
,t D, t�1
..
will
assume
that
ifare
are
in period
t, then
we have
observed
. . ,. .but
assume
thatthat
if
we
forecasting
in period
t,we
then
we like
have
Dt, Dt21
twe
will
assume
that
if D
wet11
are
forecasting
in period t, then we have observed Dt, Dt�1, . . .
.
but
observed
D.t�1
havehave
not not
observed
butDefine
have not
observed
ast the
for period
t in t-τ
period
t � demand
1. That is,
is the t,forecast
Ftt-τ,
as forecast
theDforecast
made
in period
for the
in itperiod
where
t�1.made
as
the
period
t in observed
period
t �D
That
is,
it
is
the
forecast
F.t For
,
D
,
.
.
.
,
but
before
made
at the
end
offorecast
period made
t � 1offor
after
F
.t 1.
That
is,
is
the
forecast
τ =Define
1,2,…
the
special
case
τ=
1,having
define
Ft = Ft-1,
t�1
t�2 t
, Dt�2
, .Done-step-ahead
.t-1,
.,D
but
before
made
at period
theDtend
of period
tdemand
� we
1 after
havingt, observed
Dt�1
. Hence,
forthe
now,
are
thathaving
all forecasts
are
observing
but
made in
t —1
for
in assuming
period
after
observed
t-2 …,
forDtmade
now,
assuming
all
forecasts
areforecasts
one-step-ahead
observing
Dt. Hence,
forecasts;
that
is, they are
for
the
inthat
the
period.
. Forwe
theare
timedemand
being we
willnext
assume
thatMultiple-step-ahead
all
are one
before having
observed
forecasts;
that
is, discussed
theythat
are made
the
demand
in the
next period.
Multiple-step-ahead
forecasts
be
a for
later
section.
step aheadwill
forecasts;
is,inthey
are
made
for the
demand
in the next
period. Multiple
forecasts
will
bethat
discussed
a later
section.
note
awill
timebein
series
forecast
is obtained by applying some set of weights
stepFinally,
ahead
forecasts
discussed
later.
Finally,
note
that
to past
data.
That
is,aa time
Finally,
note
that
time series
series forecast
forecast is
is obtained
obtained by
by applying
applying some
some set
set of
of weights
weights
to
to past
past data.
data. That
That is,
is, q
q anDt�n
Ft � a
for some set of weights a1, a2, . . . .
a
D
Ft � n�1
for
some set of weights a1, a2, . . . .
a n t�n
n�1
Most of the time series methods discussed in this chapter are distinguished only by the
Most of
ofofthe
the
time series methods discussed in this chapter are distinguished only by
by the
the
Most
time
choice
weights.
choice of
of weights.
weights.
choice
2.6 EVALUATING FORECASTS
2.6 Evaluating
Forecasts
2.6
EVALUATING FORECASTS
Define the forecast error in period t, et, as the difference between the forecast value for
Define
the forecast
forecast
error
value for
for
Define
the
in demand
period t,for
et, as
theperiod.
difference
between the forecast forecasts,
value
that
period
and theerror
actual
that
For multiple-step-ahead
that period and the actual
that
actual demand
demand for
forthat
thatperiod.
period.For
Formultiple-step-ahead
multiple-step-aheadforecasts,
forecasts,
et � Ft�, t � Dt,
5 FFt�,
2D
Dtt,,
eett �
t2t, tt �
and for one-step-ahead forecasts,
and for
for one-step-ahead
one-step-ahead forecasts,
forecasts, e � F � D .
and
t
t
t
5 FFtt �
2D
Dtt..
eett �
Let e1, e2, . . . , en be the forecast errors observed over n periods. Two common
Let ee11,, of
over
Two deviation
common
Let
e22,forecast
. . . , enaccuracy
be the forecast
errorsn observed
periods.
Two
common
measures
during these
periods are
thenmean
absolute
measures
of the
forecast
accuracy
during
these nngiven
periods
arefollowing
deviation
measures
of
forecast
during
these
periods
are
the mean formulas:
absolute deviation
(MAD)
and
meanaccuracy
squared error
(MSE),
by the
(MAD) and
and the
the mean
mean squared
squared error
error (MSE), given
given
by the
the following
following formulas:
formulas:
(MAD)
n by
n |ei|
MAD � (1>n) a
i�1|ei|
MAD � (1>n) a
n
i�1
n e2i
MSE � (1>n) a
i�1e2i
MSE � (1>n) a
i�1
Note that the MSE is similar to the variance
of a random sample. The MAD is
Note
the
is is
similar
to the
a random
sample.
Theit MAD
is MAD
often
the
Note
that
theMSE
MSE
similar
to variance
the variance
of aerror
random
sample.
The
is
often
thethat
preferred
method
of measuring
the of
forecast
because
does
not
require
preferred
method
ofmethod
measuring
the forecast
does notit require
squaring.
often
the preferred
of measuring
theerror
forecast
erroritbecause
does
require
squaring.
Furthermore,
when
forecast
errors
are because
normally
distributed,
as isnot
generally
Furthermore,
when forecast
are normally
as
is generally
­isssumed,
an
squaring.
Furthermore,
forecast
errors aredistributed,
normally
distributed,
generally
given by
assumed,
an estimate
ofwhen
theerrors
standard
deviation
of
the forecast
error, ase, ais
given by
1.25 times
MAD.
estimate
ofanthe
standard
deviation
of the forecast
error,
se, is
given
by
assumed,
the standard
deviation
of the
forecast
error,
1.25
times
theestimate
MAD. of
e, is the
Although
the
MAD and
and the
the MSE
MSE are
are the
the two
most common
common measures
measures of
of forecast
forecast
1.25
times thethe
MAD.
Although
MAD
two most
­accuracy,
other
measures
arethe
used
well.
One
that
not
dependent
magnitude
Although
themeasures
MAD and
MSE
are One
the
two
most
common
measures
of
forecast
accuracy,
other
are
used
asaswell.
that
isis
not
dependent
onon
thethe
magnitude
of
of the
values
demand
known
asthe
themean
mean
absolute
percentage
error
(MAPE)
andof
is
accuracy,
other
measures
are
usedasas
well.
Oneabsolute
that is not
dependenterror
on the
magnitude
the
values
of of
demand
is is
known
percentage
(MAPE)
and
is
given
by the
the
formula
the
values
offormula
demand is known as the mean absolute percentage error (MAPE) and is
given
by
n
given by the formula
n |ei >Di| d � 100.
MAPE � c (1>n) a
MAPE � c (1>n) i�1
a |ei >Di| d � 100.
i�1
nah77856_ch02.indd 62
17/12/14 6:47 PM
2.6 Evaluating Forecasts 63
Example 2.1
Artel, a manufacturer of static random access memories (SRAMs), has production plants in
Austin, Texas, and Sacramento, California. The managers of these plants are asked to forecast
production yields (measured in percent) one week ahead for their plants. Based on six weekly
forecasts, the firm’s management wishes to determine which manager is more ­successful at predicting his plant’s yields. The results of their predictions are given in the ­following spreadsheet.
Week
P1
O1
uE1u E1^2
uE1yO1u
P2
O2
uE2u E2^2
uE2yO2u
1
92
88
4
16
0.04545
96
91
5
25
0.05495
2
87
88
1
1
0.01136
89
89
0
0
0
3
95
97
2
4
0.02062
92
90
2
4
0.02222
4
90
83
7
49
0.08434
93
90
3
9
0.03333
5
88
91
3
9
0.03297
90
86
4
16
0.04651
6
93
93
0
0
0
85
89
4
16
0.04494
Cell Formulas
Cell
Formula
Copied to
D2
E2
5ABS(B2-C2)
5ABS(D2/C2)
D3:D7
E3:E7
(Similar formulas and copies for cells H2 and I2)
B10 5AVERAGE(D2:D7)
B11 5AVERAGE(I2:I7)
B13 5AVERAGE(E2:E7)
B14 5AVERAGE(J2:J7)
B16 5AVERAGE(F2:F7)
B17 5AVERAGE(K2:K7)
Interpret P1 as the forecast made by the manager of plant 1 at the beginning of each week,
O1 as the yield observed at the end of each week in plant 1, and E1 as the difference between
the predicted and the observed yields. The same definitions apply to plant 2.
Let us compare the performance of these managers using the three measures MAD, MSE, and
MAPE as defined previously. To compute the MAD we simply average the observed ­absolute errors:
MAD1 5 17y6 5 2.83
MAD2 5 18y6 5 3.00.
Based on the MADs, the first manager has a slight edge. To compute the MSE in each case,
square the observed errors and average the results to obtain
MSE1 5 79y6 5 13.17
MSE2 5 70y6 5 11.67.
The second manager’s forecasts have a lower MSE than the first, even though the MADs go
the other way. Why the switch? The reason that the first manager now looks worse is that the
MSE is more sensitive to one large error than is the MAD. Notice that the largest observed error
of 7 was incurred by manager 1.
Let us now compare their performances based on the MAPE. To compute the MAPE we
­average the ratios of the errors and the observed yields:
MAPE1 5 .0325
MAPE2 5 .0336.
Using the MAD or the MAPE, the first manager has a very slight edge, but using the MSE,
the second manager looks better. The forecasting abilities of the two managers would seem
to be very similar. Who is declared the “winner” depends on which method of evaluation
management chooses.
nah77856_ch02.indd 63
17/12/14 6:47 PM
64 Chapter Two Forecasting
Figure 2–3
Forecast error (ei)
Forecast errors over time
10
9
8
7
6
5
4
3
2
1
0
–1
–2
–3
–4
–5
–6
–7
–8
–9
– 10
1
3
5
Random
7
9
Time
11
13
15
17
Biased
A desirable property of forecasts is that they should be unbiased. Mathematically,
that means that E(ei) 5 0. One way of tracking a forecast method is to graph the values of the forecast error ei over time. If the method is unbiased, forecast errors should
fluctuate randomly above and below zero. An example is presented in Figure 2–3.
An alternative to a graphical method is to compute the cumulative sum of the forecast
errors, Sei. If the value of this sum deviates too far from zero (either above or below), it is
an indication that the forecasting method is biased. At the end of this chapter we d­ iscuss a
smoothing technique that also can be used to signal a bias in the forecasts. S
­ tatistical control charts also are used to identify unusu­ally large values of the forecast error. (Statistical
control charts are discussed in Chap­ter 12.)
Problems for Section 2.6
10. In Dave Pelz’s Short Game Bible, the author attempts to characterize the
skill of a golfer from a specific distance in terms of the ratio of the error
in the shot and the intended shot distance. For example, a five iron that is
hit 175 yards and is 20 yards off target would have an accuracy rating of
20y175 5 .114, while a sand wedge hit 60 yards that is 10 yards off target
would have an ­accuracy rating of 10y60 5 .1667 (the lower the rating, the
better). To what e­valuation method discussed in this section is this most
­similar? Why does this evaluation method make more sense in golf than
­absolute or squared ­errors?
nah77856_ch02.indd 64
17/12/14 6:47 PM
Snapshot Application
Pfizer Bets Big on Forecasts of
Drug Sales
One of Pfizer’s most successful products is Lipitor.
Lipitor, the first statin for reducing blood cholesterol,
generated annual gross sales in excess of $12 billion.
Not only was Lipitor the best selling statin, it was the
most profitable pharmaceutical ever sold until its patent expired in November 2011. At that point, generic
versions of the drug that cost less entered the market, and Lipitor’s sales dropped dramatically. Pfizer
was forced to lower prices to compete with generics and saw their profit margin sink considerably.
To counteract this loss of a valuable revenue stream,
a common strategy for drug ­companies is something
called “evergreening.” Evergreening means that the
company plans on bringing a new and more effective
drug to the market to counter the anticipated losses
when the patent runs out. In Pfizer’s case, this drug
was torcetrapib. Torcetrapib not only decreased LDL
(bad cholesterol), but increased the levels of HDL (good
c­ holesterol). In order to make the transition as smooth
as possible, Pfizer would have to have sufficient production capacity on line in 2011. Pfizer began production of torcetrapib as early as 2005 at a new $90 million
plant in Loughberg, Ireland. This was reportedly a state
of the art facility.
Unfortunately, things did not go as planned for
Pfizer. After investing $800 million in the development
of torcetrapib, Pfizer informed the United States Food
and Drug ­Administration that it was suspending Phase
3 clinical trials. It was at that point that Pfizer discovered serious side effects with the drug. They had to
make the painful decision to discontinue their efforts
after investing nearly a billion dollars. To Pfizer’s credit,
it is rare that a drug must be pulled from the market as
late as Phase 3. Still, Pfizer’s failure to accurately forecast the outcomes of the trials resulted in serious losses.
Source: Bala et al. (2011). “Competition, Capacity, and
‘Evergreening’”. Unpublished manuscript. Indian School of
Business.
11. A forecasting method used to predict can opener sales applies the follow­ing set of
weights to the last five periods of data: .1, .1, .2, .2, .4 (with .4 being applied to the
most recent observation). Observed values of can opener sales are
Period:
Observation:
1
18
2
22
3
26
4
33
5
14
6
28
7
30
8
52
Determine the following:
a. The one-step-ahead forecast for period 9.
b. The one-step-ahead forecast that was made for period 6.
12. A simple forecasting method for weekly sales of flash drives used by a local
­computer dealer is to form the average of the two most recent sales figures.
­Suppose sales for the drives for the past 12 weeks were
Week:
Sales:
1
86
2
75
3
72
4
83
5
132
6
65
7
110
8
90
9
67
10
92
11
98
12
73
a. Determine the one-step-ahead forecasts made for periods 3 through 12 using
this method.
b. Determine the forecast errors for these periods.
c. Compute the MAD, the MSE, and the MAPE based on the forecast errors
­computed in part (b).
nah77856_ch02.indd 65
17/12/14 6:47 PM
66 Chapter Two Forecasting
13. Two forecasting methods have been used to evaluate the same economic time s­ eries.
The results are
Forecast from
Forecast fromRealized Value
Method 1Method 2
of the Series
223
289
430
134
190
550
210
320
390
112
150
490
256
340
375
110
225
525
Compare the effectiveness of these methods by computing the MSE, the MAD,
and the MAPE. Do each of the measures of forecasting accuracy indicate that the
same forecasting technique is best? If not, why?
14. What does the term biased mean in reference to a particular forecasting technique?
15. What is the estimate of the standard deviation of forecast error obtained from the
data in Problem 12?
2.7 Methods for Forecasting Stationary Series
In this section we will discuss two popular techniques, moving averages and e­ xponential
smoothing, for forecasting stationary time series. A stationary time series is one in which
each observation can be represented by a constant plus a random ­fluctuation. In symbols,
Dt 5 m 1 et,
nah77856_ch02.qxd
12/31/07
4:20
where m is an unknown constant corresponding to the mean of the series and et is a
­ran­dom error with mean zero and variance s2.
The methods we consider in this section are more precisely known as single or
­simple exponential smoothing and single or simple moving averages. In addition, sinPM Page 65
gle moving averages also include weighted moving averages, which we do not discuss.
For convenience, we will not use the modifiers single and simple in what follows. The
meaning of the terms will be clear from the context.
Moving Averages
A simple but popular forecasting method is the method
of moving
averages.
ASeries
moving
2.7 Methods
for Forecasting
Stationary
65
average of order N is simply the arithmetic average of the most recent N ­observations.
, the
For the
the time
timebeing
beingwewe
restrict
attention
to one-step-ahead
forecasts.
FtforeFor
restrict
attention
to one-step-ahead
forecasts.
ThenThen
Ft, the
­fcast
orecast
made
in
period
t
2
1
for
period
t,
is
given
by
made in period t � 1 for period t, is given by
Ft � (1>N) a Di � (1>N)(Dt�1 � Dt�2 � Á � Dt�N).
t�1
i�t�N
In words,
words, this
of the
the N
N most
most recent
recent observations
observations is
used as
as the ­
the fforecast
In
this says
says that
that the
the mean
mean of
is used
orecast
for
the
next
period.
We
will
use
the
notation
MA(N)
for
N-period
moving
averages.
for the next period. We will use the notation MA(N) for N-period ­moving ­averages.
Example 2.2
Quarterly data for the failures of certain aircraft engines at a local military base during the last two
years are 200, 250, 175, 186, 225, 285, 305, 190. Both ­
Both tthree-quarter
six-quarter m
moving
hree-quarter and six-quarter ­
oving
of eengine
­aaverages
verages are used to forecast the numbers of ­
ngine failures. Determine the one-step-ahead
forecasts for periods 4 through 8 using three-period moving averages, and the one-step-ahead
forecasts for periods 7 and 8 using six-period moving averages.
Solution
The three-period moving-average forecast for period 4 is obtained by averaging the first three
data points.
F4 � (13)(200 � 250 � 175) � 208.
nah77856_ch02.indd 66
The three-period moving-average forecast for period 5 is
17/12/14 6:47 PM
2.7 Methods for Forecasting Stationary Series 67
forecasts for periods 4 through 8 using three-period moving averages, and the one-step-ahead
forecasts for periods 7 and 8 using six-period moving averages.
Solution
The three-period moving-average forecast for period 4 is obtained by averaging the first three
data points.
F4 5 (1y3)(200 1 250 1 175) 5 208.
The three-period moving-average forecast for period 5 is
F5 5 (1y3)(250 1 175 1 186) 5 204.
The six-period moving-average forecast for period 7 is
F7 5 (1y6)(200 1 250 1 175 1 186 1 225 1 285) 5 220.
Other forecasts are computed in a similar fashion. Arranging the forecasts and the associated
forecast errors in a spreadsheet, we obtain
Engine
Quarter
FailuresMA(3)ErrorMA(6)Error
1
2
3
4
5
6
7
8
200
250
175
186
208.33
22.33
225
203.67
221.33
285
195.33
289.67
305
232.00
273.00
220.17 284.83
190
271.67
81.67
237.67
47.67
Cell Formulas
Cell
C4
E7
D5
F8
Formula
51y3*SUM(B2:B4)
51y6*SUM(B2:B7)
5C5-B5
5E8-B8
Copied to
C5:C8
E8
D6:D8
F9
An interesting question is, how does one obtain multiple-step-ahead forecasts? For
example, suppose in Example 2.2 that we are interested in using MA(3) in period 3
to forecast for period 6. Because the moving-average method is based on the assumption that the demand series is stationary, the forecast made in period 3 for any ­future
nah77856_ch02.qxd 12/31/07 4:20
Page
periodPM will
be 66
the same. That is, the multiple-step-ahead and the one-step-ahead
forecasts are identical (although the one-step-ahead forecast will generally be more
accurate). Hence, the MA(3) forecast made in period 3 for p­ eriod 6 is 208. In fact,
the MA(3) forecast made in period 3 for any period beyond period 3 is 208 as well.
An apparent disadvantage of the moving-average technique is that one must
­re­compute the average of the last N observations each time a new demand observation
66 Chapter Two Forecasting
­becomes available. For large N this could be tedious. However, recalculation of the full
N-period average is not necessary every period, since
N-period average is not necessary every period, since
a
t
Ft1 (1>N)
Di (1>N) c Dt a Di DtN d
itN1
t1
itN
Ft (1>N)[Dt DtN]
This means that for one-step-ahead forecasting, we need only compute the difference
between the most recent demand and the demand N periods old in order to update the
forecast. However, we still need to keep track of all N past observations. Why?
nah77856_ch02.indd 67
Moving Average Lags behind the Trend
Consider a demand process in which there is a definite trend. For example, suppose that
the observed demand is 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24. Consider the one-step17/12/14
ahead MA(3) and MA(6) forecasts for this series.
6:47 PM
68 Chapter Two Forecasting
This means that for one-step-ahead forecasting, we need only compute the ­difference
between the most recent demand and the demand N periods old in order to update the
forecast. However, we still need to keep track of all N past observations. Why?
Moving Average Lags behind the Trend
Consider a demand process in which there is a definite trend. For example, suppose that
the observed demand is 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24. Consider the one-stepahead MA(3) and MA(6) forecasts for this series.
PeriodDemandMA(3)MA(6)
1 2 2 4 3 6 4 8 4
5
10 6
6
12 8
7
14
10 7
8
16
12 9
9
18
14
11
10
20
16
13
11
22
18
15
12
24
20
17
The demand and the forecasts for the respective periods are pictured in Figure 2–4.
Notice that both the MA(3) and the MA(6) forecasts lag behind the trend. Furthermore,
MA(6) has a greater lag. This implies that the use of simple moving averages is not an
appropriate forecasting method when there is a trend in the series.
Figure 2–4
Demand
24
Moving-average
forecasts lag behind
a trend
22
20
MA(3)
18
MA(6)
Demand
16
14
12
10
8
6
4
2
1
2
3
4
5
6
7
8
9
10
11
12
Time
nah77856_ch02.indd 68
17/12/14 6:47 PM
2.7 Methods for Forecasting Stationary Series 69
Problems on Moving Averages
Problems 16 through 21 are based on the following data. Observations of the demand
for a certain part stocked at a parts supply depot during the calendar year 2013 were
MonthDemandMonthDemand
January
February
March
April
May
June
89
57
144
221
177
280
July
August
September
October
November
December
223
286
212
275
188
312
16. Determine the one-step-ahead forecasts for the demand for January 2014 using
­3‑, 6-, and 12-month moving averages.
17. Using a four-month moving average, determine the one-step-ahead forecasts for
July through December 2013.
18. Using a four-month moving average, determine the two-step-ahead forecast for July
through December 2013. (Hint: The two-step-ahead forecast for July is based on the
observed demands in February through May.)
19. Compute the MAD for the forecasts obtained in Problems 17 and 18. Which method
gave better results? Based on forecasting theory, which method should have given
better results?
20. Compute the one-step-ahead three-month and six-month moving-average f­ orecasts
for July through December. What effect does increasing N from 3 to 6 have on the
forecasts?
21. What would an MA(1) forecasting method mean? Compare the accuracy of MA(1)
and MA(4) forecasts for July through December 2013.
Exponential Smoothing
Another very popular forecasting method for stationary time series is exponential
smoothing. The current forecast is the weighted average of the last forecast and the
current value of demand. That is,
New forecast 5 a(Current observation of demand) 1 (1 2 a)(Last forecast).
In symbols,
Ft 5 aDt21 1 (1 2 a)Ft21,
where 0 , a # 1 is the smoothing constant, which determines the relative weight
placed on the current observation of demand. Interpret (1 2 a) as the weight placed
on past observations of demand. By a simple rearrangement of terms, the exponential
smoothing equation for Ft can be written
Ft 5 Ft21 2 a(Ft21 2 Dt21)
5 Ft21 2 aet21.
nah77856_ch02.indd 69
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 68
68
Chapter Two
Forecasting
70 Chapter
Two Forecasting
Written this way, we see that exponential smoothing can be interpreted as follows:
the forecast in any period t is the forecast in period t � 1 minus some fraction of the
observed forecast error in period t � 1. Notice that if we forecast high in period t � 1,
et�1 is positive and the adjustment is to decrease the forecast. Similarly, if we forecast
low in period t � 1, the error is negative, and the adjustment is to increase the current
forecast.
As before, Ft is the one-step-ahead forecast for period t made in period t � 1. Notice
that since
Ft�1 � Dt�2 � (1 � )Ft�2,
we can substitute above to obtain
Ft � Dt�1 � (1 � )Dt�2 � (1 � )2Ft�2.
We can now substitute for Ft�2 in the same fashion. If we continue in this way, we
obtain the infinite expansion for Ft,
Ft � a (1 � )iDt�i�1 � a aiDt�i�1,
q
q
i�0
i�0
where the weights are a0 � a1 � a2 � � � � � ai � (1 � )i, and
i
i
a ai � a (1 � ) � a (1 � ) � � 1>[1 � (1 � )] � 1.
q
q
q
i�0
i�0
i�0
Hence, exponential smoothing applies a declining set of weights to all past data. The
weights are graphed as a function of i in Figure 2–5.
In fact, we could fit the continuous exponential curve g(i) � exp(�i) to these
weights, which is why the method is called exponential smoothing. The smoothing
constant plays essentially the same role here as the value of N does in moving
FIGURE 2–5
Weights in exponential smoothing
Weight applied to data i periods old
0.1
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
nah77856_ch02.indd 70
· · ·
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
i
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 69
2.7 Methods for
for Forecasting
Forecasting Stationary
Stationary Series 71
Series 69
2.7 Methods
averages. If is large, more weight is placed on the current observation of demand and
less weight on past observations, which results in forecasts that will react quickly
to changes in the demand pattern but may have much greater variation from period to
period. If is small, then more weight is placed on past data and the forecasts are
more stable.
When using an automatic forecasting technique to predict demand for a production
application, stable forecasts (that is, forecasts that do not vary a great deal from period
to period) are very desirable. Demand forecasts are used as the starting point for production planning and scheduling. Substantial revision in these forecasts can wreak
havoc with employee work schedules, component bills of materials, and external purchase orders. For this reason, a value of between .1 and .2 is generally recommended
for production applications. (See, for example, Brown, 1962.)
Multiple-step-ahead forecasts are handled the same way for simple exponential
smoothing as for moving averages; that is, the one-step-ahead and the multiple-stepahead forecasts are the same.
Example 2.3
Consider Example 2.2 in which moving averages were used to predict aircraft engine
failures. The observed numbers of failures over a two-year period were 200, 250, 175,
186, 225, 285, 305, 190. We will now forecast using exponential smoothing. In order to get
the method started, let us assume that the forecast for period 1 was 200. Suppose that � .1.
The one-step-ahead forecast for period 2 is
F2 � D1 � (1 � )F1 � (.1)(200) � (.9)(200) � 200.
Similarly,
F3 � D2 � (1 � )F2 � (.1)(250) � (.9)(200) � 205.
Other one-step-ahead forecasts are computed in the same fashion. The observed numbers of
failures and the one-step-ahead forecasts for each quarter are the following:
Quarter
1
2
3
4
5
6
7
8
Failures
200
250
175
186
225
285
305
190
Forecast
200 (by assumption)
200
205
202
201
203
211
220
Notice the effect of the smoothing constant. Although the original series shows high variance, the forecasts are quite stable. Repeat the calculations with a value of � .4. There will
be much greater variation in the forecasts.
Because exponential smoothing requires that at each stage we have the previous
forecast, it is not obvious how to get the method started. We could assume that the
initial forecast is equal to the initial value of demand, as we did in Example 2.3.
However, this approach has a serious drawback. Exponential smoothing puts substantial weight on past observations, so the initial value of demand will have an
unreasonably large effect on early forecasts. This problem can be overcome by
nah77856_ch02.indd 71
17/12/14 6:47 PM
72 Chapter Two Forecasting
Figure 2–6
Exponential ­smoothing
for ­different values of
alpha
100
90
80
Demand
70
60
50
40
30
20
10
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Time
Demand
ES(.4)
ES(.1)
a­ llowing the process to evolve for a reasonable number of periods (say 10 or more) and
using the arithmetic average of the demand during those periods as the initial forecast.
In order to appreciate the effect of the smoothing constant, we have graphed a particularly
erratic series in Figure 2–6, along with the resulting forecasts using values of a 5 .1 and
a 5 .4. Notice that for a 5 .1 the predicted value of demand results in a r­ elatively smooth
pattern, whereas for a 5 .4, the predicted value exhibits significantly greater variation.
­Although smoothing with the larger value of a does a better job of tracking the series, the
stability afforded by a smaller smoothing constant is very desirable for planning purposes.
Example 2.3
(continued)
Consider again the problem of forecasting aircraft engine failures. Suppose that we were
i­nterested in comparing the performance of MA(3) with the exponential smoothing forecasts
­obtained earlier (ES(.1)). The first period for which we have a forecast using MA(3) is period 4,
so we will make the comparison for periods 4 through 8 only.
Quarter
1
2
3
4
5
6
7
8
Failures
200
250
175
186
225
285
305
190
Forecast
200.00
200.00 (by assumption)
205.00
202.00
200.40
202.86
211.07
220.47
Cell Formulas
Cell
C3
Formula
5SUM(B$12*B2,(1-B$12)*C2)
Copied to
C4:C9
The arithmetic average of the absolute errors, the MAD, is 57.6 for the three-period moving
average and 49.2 for exponential smoothing. The respective values of the MSE are 4,215.6 and
3,458.4. Based on this comparison only, one might ­conclude that exponential smoothing is a
superior method for this series. This is not necessarily true, however.
nah77856_ch02.indd 72
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 71
2.7
Methods for
71
2.7 Methods
for Forecasting
Forecasting Stationary
Stationary Series
Series 73
In Example 2.3 we compared exponential smoothing with � .1 and a moving
average with N � 3. How do we know that those parameter settings are consistent?
The MA(3) forecasts exhibit much greater variability than the ES(.1) forecasts do,
suggesting that � .1 and N � 3 are not consistent values of these parameters.
Determining consistent values of and N can be done in two ways. One is to equate the
average age of the data used in making the forecast. A moving-average forecast consists
of equal weights of 1N applied to the last N observations. Multiplying the weight placed
on each observation by its “age,” we get the average age of data for moving averages as
Average age � (1N)(1 � 2 � 3 � � � � � N ) � (1N )(N )(N � 1)2
� (N � 1)2.
For exponential smoothing, the weight applied to data i periods old is (1 � )i�1.
Assume that we have an infinite history of past observations of demand. Hence, the
average age of data in an exponential smoothing forecast is
Average age � a i(1 � )i�1 � 1>.
q
i�1
We omit the details of this calculation.
Equating the average age of data for the two methods, we obtain
N�1 1
� ,
2
which is equivalent to
� 2(N � 1)
or
N�
2�
.
Hence, we see that we would have needed a value of N � 19 for � .1 or a value
of � .5 for N � 3 in order for the methods to be consistent in the sense of average
age of data.
In Appendix 2–A of this chapter, we derive the mean and variance of the forecast
error for both moving averages and exponential smoothing in terms of the variance of
each individual observation, assuming that the underlying demand process is stationary. We show that both methods are unbiased; that is, the expected value of the forecast
error is zero. Furthermore, by equating the expressions for the variances of the forecast
error, one obtains the same relationship between and N as by equating the average
age of data. This means that if both exponential smoothing and moving averages are
used to predict the same stationary demand pattern, forecast errors are normally distributed, and � 2(N � 1), then both methods will have exactly the same distribution
of forecast errors. (However, this does not mean that the forecasts obtained by the two
methods are the same.)
Multiple-Step-Ahead Forecasts
Thus far, we have talked only about one-step-ahead forecasts. That is, we have assumed
that a forecast in period t is for the demand in period t � 1. However, there are cases
where we are interested in making a forecast for more than one step ahead. For example, a retailer planning for the Christmas season might need to make a forecast for
December sales in June in order to have enough time to prepare. Since the underlying
model assumed for both moving averages and exponential smoothing is stationary
(i.e., not changing in time), the one-step-ahead and multiple-step-ahead forecasts for
nah77856_ch02.indd 73
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 72
nah77856
72
Chapter Two
Forecasting
74 Chapter
Two Forecasting
moving averages and exponential smoothing are the same. That is, a forecast made in
June for July sales is the same as a forecast made in June for December sales. (In the
case of the retailer, the assumption of stationarity would probably be wrong, since
December sales would likely be greater than a typical month’s sales. That would suggest that these methods would not be appropriate in this case.)
Comparison of Exponential Smoothing and Moving Averages
There are several similarities and several differences between exponential smoothing
and moving averages.
Similarities
1. Both methods are derived with the assumption that the underlying demand process
is stationary (that is, can be represented by a constant plus a random fluctuation with zero
mean). However, we should keep in mind that although the methods are appropriate for
stationary time series, we don’t necessarily believe that the series are stationary forever.
By adjusting the values of N and we can make the two methods more or less responsive to shifts in the underlying pattern of the data.
2. Both methods depend on the specification of a single parameter. For moving
averages the parameter is N, the number of periods in the moving average, and for
exponential smoothing the parameter is , the smoothing constant. Small values of N
or large values of result in forecasts that put greater weight on current data, and large
values of N and small values of put greater weight on past data. Small N and large may be more responsive to changes in the demand process, but will result in forecast
errors with higher variance.
3. Both methods will lag behind a trend if one exists.
4. When � 2(N � 1), both methods have the same distribution of forecast error.
This means that they should have roughly the same level of accuracy, but it does not
mean that they will give the same forecasts.
Differences
1. The exponential smoothing forecast is a weighted average of all past data points
(as long as the smoothing constant is strictly less than 1). The moving-average forecast is
a weighted average of only the last N periods of data. This can be an important advantage
for moving averages. An outlier (an observation that is not representative of the sample
population) is washed out of the moving-average forecast after N periods, but remains
forever in the exponential smoothing forecast.
2. In order to use moving averages, one must save all N past data points. In order to
use exponential smoothing, one need only save the last forecast. This is the most
significant advantage of the exponential smoothing method and one reason for its popularity in practice. In order to appreciate the consequence of this difference, consider a
system in which the demand for 300,000 inventory items is forecasted each month
using a 12-month moving average. The forecasting module alone requires saving
300,000 � 12 � 3,600,000 pieces of information. If exponential smoothing were used,
only 300,000 pieces of information need to be saved. This issue is less important today
than it has been, as the cost of information storage has decreased enormously in recent
years. However, it is still easier to manage a system that requires less data. It is primarily for this reason that exponential smoothing appears to be more popular than
moving averages for production-planning applications.
nah77856_ch02.indd 74
17/12/14 6:47 PM
nah77856_ch02.qxd
1/29/08
5:12 PM
Page 73
2.7 Methods for
for Forecasting
Forecasting Stationary
Stationary Series 75
Series 73
2.7 Methods
Problems for Section 2.7
22. Handy, Inc., produces a solar-powered electronic calculator that has experienced
the following monthly sales history for the first four months of the year, in
thousands of units:
January
February
23.
24.
25.
26.
27.
nah77856_ch02.indd 75
23.3
72.3
March
April
30.3
15.5
a. If the forecast for January was 25, determine the one-step-ahead forecasts for
February through May using exponential smoothing with a smoothing constant of � .15.
b. Repeat the calculation in part (a) for a value of � .40. What difference in the
forecasts do you observe?
c. Compute the MSEs for the forecasts you obtained in parts (a) and (b) for
February through April. Which value of gave more accurate forecasts, based
on the MSE?
Compare and contrast exponential smoothing when is small (near zero) and
when is large (near 1).
Observed weekly sales of ball peen hammers at the town hardware store over an
eight-week period have been 14, 9, 30, 22, 34, 12, 19, 23.
a. Suppose that three-week moving averages are used to forecast sales. Determine
the one-step-ahead forecasts for weeks 4 through 8.
b. Suppose that exponential smoothing is used with a smoothing constant
of � .15. Find the exponential smoothing forecasts for weeks 4 through 8.
[To get the method started, use the same forecast for week 4 as you used in
part (a).]
c. Based on the MAD, which method did better?
d. What is the exponential smoothing forecast made at the end of week 6 for the
sales in week 12?
Determine the following:
a. The value of consistent with N � 6 in moving averages.
b. The value of N consistent with � .05.
c. The value of that results in a variance of forecast error, 2e, 10 percent higher
than the variance of each observation, 2 (refer to the formulas derived in
Appendix 2–A).
Referring to the data in Problem 22, what is the exponential smoothing forecast
made at the end of March for the sales in July? Assume � .15.
For the data for Problems 16 through 21, use the arithmetic average of the first six
months of data as a baseline to initialize the exponential smoothing.
a. Determine the one-step-ahead exponential smoothing forecasts for August
through December, assuming � .20.
b. Compare the accuracy of the forecasts obtained in part (a) with the one-stepahead six-month moving-average forecasts determined in Problem 20.
c. Comment on the reasons for the result you obtained in part (b).
17/12/14 6:47 PM
Snapshot Application
Sport obermeyer slashes costs with
improved forecasting1
Sport Obermeyer is a leading supplier in the U.S. ­fashion
ski apparel market. The firm, was founded in 1950 by
engineer/ski instructor Klaus Obermeyer. Virtually all
the firm’s offerings are ­redesigned annually to incorporate changes in style, fabrics, and colors. For more
than 50 years, the firm was able to ­successfully meet
demands by producing during the summer months
­
after receiving firm orders from customers.
During the 1990s, things changed and problems
­developed. First, volumes increased. There was insufficient capacity among suppliers to produce the required
volume in the summer. Second, the firm developed a
complex global supply chain strategy (see Section 6.10)
to reduce costs. A parka sold in the United States might
be sewn in China from fabrics and parts from Japan,
South Korea, and ­
Germany. Together these changes
lengthened the production lead time, thus requiring
the firm to commit to production before orders were
placed by customers.
The firm undertook several “quick response” initiatives to reduce lead times. These included encouraging
some customers to place orders earlier, locating raw
materials near the Far East production facility, and instituting an air freight system to expedite delivery from
the Far East to its Denver distribution center. Even with
these changes in place, the problem of stockouts and
markdowns due to oversupply were not solved. The
company still had to commit about half the production based on forecasts. In the fashion i­ndustry, there
is often no statistical history on which to base forecasts,
and forecast errors can be huge. Products that outsell
original forecasts by a factor of 2 or undersell original
forecasts by a factor of 10 are common.
Sport Obermeyer needed some help with forecasting to avoid expensive miscalculations. The customary
procedure was to base the forecasts on a consensus of
members of the buying committee. The problem with
1
This application is based on the work of a team from the
Wharton School and the Harvard Business School. The ­results
are reported in Fisher et al. (1994).
nah77856_ch02.indd 76
consensus forecasting is that the dominant personalities in a group carry more weight. A forecast obtained
in this way might represent only the opinion of one
­person. To overcome this problem, the research team
(Fisher et al., 1994) recommended that members of the
committee supply ­individual forecasts.
The dispersion among individual forecasts turned
out to be a reliable indicator of forecast accuracy.
When committee members’ forecasts were close, forecasts were more accurate. This provided a mechanism
for signaling the products whose sales were likely to
be poorly forecast. This did not solve the problem of
poorly forecast items, but it allowed the firm to commit
first to production of items whose forecasts were likely
to be accurate. By the time production had to begin on
the problem items, information on early sales patterns
would be available.
The team noticed that retailers were remarkably
similar. That meant that even if only the first 20 percent of orders for a product were in, that information
could dramatically improve forecasts. Production plans
for these “trouble” items could now be committed with
greater confidence. In this way, the firm could separate
products into two categories: reactive and nonreactive.
The nonreactive items are those for which the forecast
is likely to be accurate. These are produced early in the
season. The reactive items are those whose forecasts
are updated later in the season from early sales figures.
The firm’s experience was that stockout and markdown
costs were reduced from 10.2 percent of sales to 1.8 percent of sales on items that could be produced reactively.
Sport Obermeyer was able to produce 30 percent of its
season’s volume reactively and experienced a cost reduction of about 5 percent of sales.
What are the lessons here? One is that even in cases
where there is no statistical history, statistical methodology can be successfully applied to improve forecasting
accuracy. Another is not to assume that things should
be done a certain way. Sport Obermeyer assumed that
consensus forecasting was the best ­approach. In fact,
by requiring the buying committee to reach a consensus, valuable information was being ­ignored. The
­differences among individual forecasts proved to be
­important.
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 75
2.8 Trend-Based Methods 77
Methods 75
2.8 Trend-Based
2.8
TREND-BASED METHODS
Both exponential smoothing and moving-average forecasts will lag behind a trend if one
exists. We will consider two forecasting methods that specifically account for a trend in
the data: regression analysis and Holt’s method. Regression analysis is a method that fits
a straight line to a set of data. Holt’s method is a type of double exponential smoothing
that allows for simultaneous smoothing on the series and on the trend.
Regression Analysis
Let (x1, y1), (x2, y2), . . . , (xn, yn) be n paired data points for the two variables X and Y.
Assume that yi is the observed value of Y when xi is the observed value of X. Refer to
Y as the dependent variable and X as the independent variable. We believe that a relationship exists between X and Y that can be represented by the straight line
Ŷ � a � bX.
Interpret Ŷ as the predicted value of Y. The goal is to find the values of a and b so
that the line Ŷ � a � bX gives the best fit of the data. The values of a and b are chosen
so that the sum of the squared distances between the regression line and the data points
is minimized (see Figure 2–7). In Appendix 2–B we derive the optimal values of a and
b in terms of the given data.
When applying regression analysis to the forecasting problem, the independent variable often corresponds to time and the dependent variable to the series to be forecast.
Assume that D1, D2, . . . , Dn are the values of the demand at times 1, 2, . . . , n. Then it
is shown in Appendix 2–B that the optimal values of a and b are given by
b�
Sxy
Sxx
FIGURE 2–7
Demand
An example of a
regression line
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
2
3
4
5
6
7
8
9
10
11
12
Time
nah77856_ch02.indd 77
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 76
76 Chapter Two Forecasting
76
Chapter Two
Forecasting
78 Chapter
Two Forecasting
and
a � D � b(n � 1)>2,
and
where
a � D � b(n � 1)>2,
Sxy � n a iDi �
n(n � 1) n
a Di,
2
n
n
i�1
n(n � 1) i�1
,
Sxy � n 2a iDi �
an2D
n (n � 1)(2n �2 1) i�1
(ni � 1)2
,
�
Sxx � i�1
6
2
2 4
2
n (n � 1)(2n � 1) n (n � 1)
Sxx �
�
,
4 during periods 1, 2, . . . , n.
and D is the arithmetic average of the6observed demands
where
Example 2.4
Example 2.4
n
and D is the arithmetic average of the observed demands during periods 1, 2, . . . , n.
We will apply regression analysis to the problem, treated in Examples 2.2 and 2.3, of predicting
aircraft engine failures. Recall that the demand for aircraft engines during the last eight quarters
We will
apply
to the
problem,
treated in
Examples
predicting
was
200,
250,regression
175, 186,analysis
225, 285,
305,
190. Suppose
that
we use 2.2
the and
first2.3,
five of
periods
as a
aircraft engine
failures.
Recallour
thatregression
the demand
for aircraftThen
engines during the last eight quarters
baseline
in order
to estimate
parameters.
was 200, 250, 175, 186, 225, 285, 305, 190. Suppose that we use the first five periods as a
Sxyto�estimate
5[200 �our
(250)(2)
� (175)(3)
� (186)(4)
baseline in order
regression
parameters.
Then� (225)(5)]
� [(5)(6)2][200 � 250 � 175 � 186 � 225]
Sxy � 5[200 � (250)(2) � (175)(3) � (186)(4) � (225)(5)]
� �70,
� [(5)(6)2][200 � 250 � 175 � 186 � 225]
� (25)(36)4 � 50.
Sxx �
� (25)(6)(11)6
�70,
Then
Sxx � (25)(6)(11)6 � (25)(36)4 � 50.
b � SxySxx � �7050 � �75,
Then
a � 207.2 � (�75)(3) � 211.4.
b � SxySxx � �7050 � �75,
It follows that the regression equation
based
on five periods
of data is
a � 207.2
� (�75)(3)
� 211.4.
D̂ � 211.4 � (75)t.
It follows that the regression equation tbased on five periods of data is
D̂t is the predicted value of demand at time t. We would use this regression equation to
D̂t � 211.4 � (75)t.
forecast from period 5 to any period beyond
period 5. For example, the forecast made in period
D̂t is
of demand
time
t. Wet.would
usethe
this
regression
equationequation
to forecast
isthe
thepredicted
predicted
value
of demand
at time
use
this regression
to
5 for
period
8 wouldvalue
be
obtained
by at
substituting
tWe
� 8would
in
regression
equation
just
given,
from period
5 to
any in
period
beyond
period
5.
example,
the
forecast
made
period
5interested
for
forecast
from
period
5 the
to any
period
beyond
period
5.�For
example,
forecast
made
in period
which
would
result
forecast
211.4
�For
(75)(8)
200.2.
Notethe
that
ifin ­
we
were
8 would
be obtained
bybe
substituting
t8,
5substituting
8 in the
just
given,
which
would
result
5
for
period
8inwould
t � 8 equation
inequation
the regression
equation
just given,
in
forecasting
period
7 obtained
for periodby
then
thisregression
regression
would
not
be appropriate.
in the
forecast
211.4
2 the
(7y5)(8)
200.2.
Note
thatusing
if we
interested
in
forecasting
in
period
7
which
would
result
in
forecast
211.4
� (75)(8)
� were
200.2.
Note
that
if we 1
were
interested
We
would
have
to repeat
the 5
entire
calculation
the data
from
periods
through
7. In
forforecasting
period
8, the
then
this regression
equation
would
be ­afor
ppropriate.
have
repeat
the
in
inserious
period
7 for period
then
thisnot
regression
equationWe
would
be to
appropriate.
fact,
one of
drawbacks
of8,using
regression
forecasting
iswould
thatnot
updating
forecasts
entire
calculation
using
the data
from
1 through
7. Indata
fact,from
one of
the serious
drawbacks
We
would
to
repeat
the isentire
calculation
using the
periods
1 through
7. In
as
new
datahave
become
available
very periods
cumbersome.
of using
for forecasting
that
updating ­
forecasts
as new data
available
is very
fact,
oneregression
of the serious
drawbacksisof
using
regression
for forecasting
is become
that updating
forecasts
cumbersome.
(Note that
Excel includes
single and ­multiple linear regression capabilities.)
as
new data become
available
is very cumbersome.
56_ch02.qxd
Problems
Problems for
for Section
Section 2.8
2.8
28. Shoreline
Park in Mountain View, California, has kept close tabs on the number of
Problems for Section
2.8
28. Shoreline
Park in Mountain View, California, has kept close tabs on the number of
12/31/07
4:21 PM
patrons using the park since its opening in January 1993. For the first six months
using
the Mountain
park sinceView,
its opening in January
1993.
For
theonfirst
months
28. patrons
Shoreline
Park
has kept
close
tabs
thesix
number
of
of operation,
theinfollowing
figuresCalifornia,
were recorded:
of
2013,
the
following
figures
were
recorded:
patrons using the park since its opening in January 1993. For the first six months
of operation, the following figures were recorded:
Page 77
Month
Number of Patrons
Month
Number of Patrons
January
April
640
Month
Number133
of Patrons
Month
Number of
Patrons
February
183
May
1,876
2.8 Trend-Based Methods 77
March
285
June
2,550
January
133
April
640
February
183
May
1,876
285
a. Draw a graph March
of these six data points.
Assume June
that January �2,550
period 1,
nah77856_ch02.indd 78
February � period 2, and so on. Using a ruler, “eyeball” the best straight-line
fit of the data. Estimate the slope and intercept from your graph.
b. Compute the exact values of the intercept a and the slope b from the regression
equations.
c. What are the forecasts obtained for July through December 2013 from the
regression equation determined in part (b)?
d. Comment on the results you obtained in part (c). Specifically, how confident
would you be about the accuracy of the forecasts that you obtained?
17/12/14 6:47 PM
2.8
Trend-Based Methods 77
a. Draw a graph of these six data points. Assume that January � period 1,
February � period 2, and so on. Using a ruler, “eyeball”
the best straight-line
2.8 Trend-Based
Methods 79
fit of the data. Estimate the slope and intercept from your graph.
b. Compute the exact values of the intercept a and the slope b from the regression
equations.
c. What are the forecasts obtained for July through December 2013 from the
regression equation determined in part (b)?
d. Comment on the results you obtained in part (c). Specifically, how confident
would you be about the accuracy of the forecasts that you obtained?
29. The Mountain View Department of Parks and Recreation must project the total use
of Shoreline Park for calendar year 2014.
a. Determine the forecast for the total number of people using the park in 2014
based on the regression equation.
b. Determine the forecast for the same quantity using a six-month moving average.
c. Draw a graph of the most likely shape of the curve describing the park’s usage by
month during the calendar year, and predict the same quantity using your graph.
Is your prediction closer to the answer you obtained for part (a) or part (b)?
Discuss your results.
Double Exponential Smoothing Using Holt’s Method
Holt’s method is a type of double exponential smoothing designed to track time series
with linear trend. The method requires the specification of two smoothing constants, and , and uses two smoothing equations: one for the value of the series (the intercept)
and one for the trend (the slope). The equations are
St � Dt � (1 � )(St�1 � Gt�1),
Gt � (St � St�1) � (1 � )Gt�1.
Interpret St as the value of the intercept at time t and Gt as the value of the slope at
time t. The first equation is very similar to that used for simple exponential smoothing.
When the most current observation of demand, Dt, becomes available, it is averaged
with the prior forecast of the current demand, which is the previous intercept, St�1, plus
1 times the previous slope, Gt�1. The second equation can be explained as follows.
Our new estimate of the intercept, St, causes us to revise our estimate of the slope to
St � St�1. This value is then averaged with the previous estimate of the slope, Gt�1.
The smoothing constants may be the same, but for most applications more stability is
given to the slope estimate (implying � ).
The -step-ahead forecast made in period t, which is denoted by Ft, t�, is given by
Ft, t� � St � Gt.
Example 2.5
Let us apply Holt’s method to the problem of developing one-step-ahead forecasts for the
aircraft engine failure data. Recall that the original series was 200, 250, 175, 186, 225, 285,
305, 190. Assume that both and are equal to .1. In order to get the method started,
we need estimates of both the intercept and the slope at time zero. Suppose that these are
S0 � 200 and G0 � 10. Then we obtain
S1 � (.1)(200) � (.9)(200 � 10) � 209.0
G
(.1)(209 200) � (.9)(10) 9.9
S21 5
(.1)(250)
1 (.9)(209 1 9.9)�
5 222.0
�
�
G2 5 (.1)(222 2 209) 1 (.9)(9.9) 5 10.2
S3 5 (.1)(175) 1 (.9)(222 1 10.2) 5 226.5
G3 5 (.1)(226.5 2 222) 1 (.9)(10.2) 5 9.6
and so on.
nah77856_ch02.indd 79
17/12/14 6:47 PM
nah77856
80 Chapter Two Forecasting
Comparing the one-step-ahead forecasts to the actual numbers of failures for periods 4
through 8, we obtain the following:
PeriodActualSG Forecast
uErroru
1
2
3
4
5
6
7
8
0.00
31.10
57.22
50.14
15.26
37.28
44.19
85.02
200
250
175
186
225
285
305
190
200.00
209.00
222.01
226.50
231.12
238.74
251.45
265.23
10.00
9.90
10.21
9.64
9.14
8.98
9.36
9.80
200.00
218.90
232.22
236.14
240.26
247.72
260.81
275.02
  0.1
  0.1
So  200
Go  10
Cell Formulas
Cell
C3
D3
E3
F3
Formula
SUM(B$12*B2,(1-B$12)*(C2+D2))
SUM(B$13*(C3-C2),(1-B$13)*(D2))
C3+D3
ABS(B3-E3)
Copied to
C4:C9
D4:D9
E4:E9
F4:F9
Averaging the numbers in the final column, we obtain a MAD of 46.4. Notice that this is
lower than that for simple exponential smoothing or moving averages. Holt’s method does better for this series because it is explicitly designed to track the trend in the data, whereas simple
exponential smoothing and moving averages are not. Note that the forecasts in the given table
are one-step-ahead forecasts. Suppose you needed to forecast the demand in period 2 for
­period 5. This forecast is F2,5 5 S2 1 (3)G2 5 222 1 (3)(10.2) 5 252.6.
The initialization problem also arises in getting Holt’s method started. The best
a­ pproach is to establish some set of initial periods as a baseline and use regression
analysis to determine estimates of the slope and intercept using the baseline data.
Both Holt’s method and regression are designed to handle series that exhibit trend.
However, with Holt’s method it is far easier to update forecasts as new observations
­become available.
More Problems for Section 2.8
30. For the data in Problem 28, use the results of the regression equation to estimate
the slope and intercept of the series at the end of June. Use these numbers as
the initial values of slope and intercept required in Holt’s method. Assume that
a 5 .15, b 5 .10 for all calculations.
a. Suppose that the actual number of visitors using the park in July was 2,150 and
the number in August was 2,660. Use Holt’s method to update the ­estimates of
the slope and intercept based on these observations.
b. What are the one-step-ahead and two-step-ahead forecasts that Holt’s method
gives for the number of park visitors in September and October?
c. What is the forecast made at the end of July for the number of park ­attendees in
December?
31. Because of serious flooding, the park was closed for most of December 1993. During that time only 53 people visited. Comment on the effect this observation would
have on predictions of future use of the park. If you were in charge of forecasting
the park’s usage, how would you deal with this data point?
nah77856_ch02.indd 80
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 79
2.9 Methods for Seasonal Series 81
32. Discuss some of the problems that could arise when using either regression analy2.9 Methods for Seasonal Series 79
sis or Holt’s method for obtaining multiple-step-ahead forecasts.
2.9
METHODS FOR SEASONAL SERIES
This section considers forecasting methods for seasonal problems. A seasonal series
is one that has a pattern that repeats every N periods for some value of N (which is at
least 3). A typical seasonal series is pictured in Figure 2–8.
We refer to the number of periods before the pattern begins to repeat as the length
of the season (N in the picture). Note that this is different from the popular usage of the
word season as a time of year. In order to use a seasonal model, one must be able to
specify the length of the season.
There are several ways to represent seasonality. The most common is to assume that
there exists a set of multipliers ct, for 1 � t � N, with the property that � ct � N. The
multiplier ct represents the average amount that the demand in the t th period of the season is above or below the overall average. For example, if c3 � 1.25 and c5 � .60, then,
on average, the demand in the third period of the season is 25 percent above the average demand and the demand in the fifth period of the season is 40 percent below the average demand. These multipliers are known as seasonal factors.
Seasonal Factors for Stationary Series
In this part of the section we present a simple method of computing seasonal factors for
a time series with seasonal variation and no trend. In the next part of this section we
consider a method that is likely to be more accurate when there is also trend. Both methods require a minimum of two seasons of data.
The method is as follows:
1. Compute the sample mean of all the data.
2. Divide each observation by the sample mean. This gives seasonal factors for each
period of observed data.
3. Average the factors for like periods within each season. That is, average all the factors corresponding to the first period of a season, all the factors corresponding to the
FIGURE 2–8
26
A seasonal demand
series
24
22
20
Demand
18
16
14
12
10
8
6
4
2
0
N
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16
Time
nah77856_ch02.indd 81
17/12/14 6:47 PM
82 Chapter Two Forecasting
second period of a season, and so on. The resulting averages are the N seasonal
­factors. They will always add to exactly N.
Example 2.6
The County Transportation Department needs to determine usage factors by day of the week
for a popular toll bridge connecting different parts of the city. In the current study, they are
considering only working days. Suppose that the numbers of cars using the bridge each working day over the past four weeks were (in thousands of cars)
Monday
Tuesday
Wednesday
Thursday
Friday
Week 1
Week 2
Week 3
Week 4
16.2
12.2
14.2
17.3
22.5
17.3
11.5
15.0
17.6
23.5
14.6
13.1
13.0
16.9
21.9
16.1
11.8
12.9
16.6
24.3
Find the seasonal factors corresponding to daily usage of the bridge.
Solution
To solve this problem we
1. Compute the arithmetic average of all of the observations (20 in this case).
2. Divide each observation by the average computed in step 1. This will give 20 factors.
3. Average factors corresponding to the same period of each season. That is, average all
­factors for Mondays, all factors for Tuesdays, and so on. This will give five seasonal factors:
one for each day of the week. Note that these factors will sum to five.
4. Forecasts for the numbers of cars using the bridge by day of the week are obtained by
­multiplying the seasonal factors computed in step 3 by the average value computed in step 1.
These steps are summarized in the spreadsheet below.
Monday
Tuesday
Wednesday
Thursday
Friday
Week 1
Week 2
Week 3
Week 4
16.20
12.20
14.20
17.30
22.50
17.30
11.50
15.00
17.60
23.50
14.60
13.10
13.00
16.90
21.90
16.10
11.80
12.90
16.60
24.30
Step 1: Compute the overall average of all of the observations
Average = 16.43
Formula = AVERAGE(B2:E6)
Step 2: Divide each observation by the Mean
Monday
Tuesday
Wednesday
Thursday
Friday
nah77856_ch02.indd 82
Week 1
Week 2
Week 3
Week 4
0.99
0.74
0.86
1.05
1.37
1.05
0.70
0.91
1.07
1.43
0.89
0.80
0.79
1.03
1.33
0.98
0.72
0.79
1.01
1.48
17/12/14 6:47 PM
2.9 Methods for Seasonal Series 83
Step 3: Average Factors corresponding to the same day of the week
Seasonal Factor
Monday
Tuesday
Wednesday
Thursday
Friday
0.98
0.74
0.84
1.04
1.40
Note that these factors sum to exactly five.
Step 4: Build the forecasts by multiplying the mean, 16.425, by the appropriate factor
Forecast
Monday
Tuesday
Wednesday
Thursday
Friday
16.05
12.15
13.78
17.10
23.05
Cell Formulas
Cell
Formula
B10
AVERAGE(B2:E6)
Copied to
B14
B2/$B$10
B15:B18
C14
C2/$C$10
C15:C18
(Similar formulas and copies for cells D14 and E14)
B22
AVERAGE(B14:E14)
B23:B26
B30
$B$10*B22
B31:B34
Note: Example 2.6
Determining the Deseasonalized Series
In cases where there is both seasonal variation and trend, a useful technique is to form
the deseasonalized series by removing the seasonal variation from the original series.
To illustrate this, consider the following simple example which consists of two seasons
of data.
Example 2.7
PeriodDemand
1
2
3
4
5
6
7
8
nah77856_ch02.indd 83
10
20
26
17
12
23
30
22
17/12/14 6:47 PM
84 Chapter Two Forecasting
Following the steps described above, the reader should satisfy himself or herself that we obtain
the following four seasonal factors in this case:
0.550
1.075
1.400
0.975
To obtain the deseasonalized series, one simply divides each observation by the appropriate
seasonal factor. For this example, that’s 10/.550, 20/1.075, etc. In this case one obtains
PeriodDemandDeseasonalized Demand
1
2
3
4
5
6
7
8
10
20
26
17
12
23
30
22
18.182
18.605
18.571
17.436
21.818
21.395
21.429
22.564
Notice that the deseasonalized demand shows a clear trend. To forecast the deseasonalized
s­ eries one could use any of the trend based methods discussed earlier in this chapter. Suppose
that we fit a simple linear regression to this data where time is the independent variable as
described in Section 2.8. Doing so one obtains the regression fit of this data as yt 5 16.91 1
0.686t. To forecast, one first applies the regression to forecast the deseasonalized series, and
then re-seasonalizes by multiplying by the appropriate factor. For example, if one wishes to
forecast for periods 9 through 12, the regression equation gives the following forecasts for the
deseasonalized series: 23.08, 23.77, 24.46, 25.14. The final forecasts are obtained by multiplying by the appropriate seasonal factors, giving the final forecasts for periods 9 through 12 as
12.70, 25.55, 34.24, and 25.51.
Problems for Section 2.9
33. Sales of walking shorts at Hugo’s Department Store in downtown Rolla appear to
exhibit a seasonal pattern. The proprietor of the store, Wally Hugo, has kept careful
records of the sales of several of his popular items, including walking shorts. During
the past two years the monthly sales of the shorts have been
Year 1Year 2Year 1Year 2
Jan.
Feb.
March
April
May
June
12
18
36
53
79
134
16
14
46
48
88
160
July
Aug.
Sep.
Oct.
Nov.
Dec.
112
90
66
45
23
21
130
83
52
49
14
26
Assuming no trend in shorts sales over the two years, obtain estimates for the
monthly seasonal factors.
34. A popular brand of tennis shoe has had the following demand history by quarters
over a three-year period.
nah77856_ch02.indd 84
17/12/14 6:47 PM
Snapshot Application
2.9 Methods for Seasonal Series 85
Nate Silver Perfectly Forecasts 2012
Presidential Election
In a stunning victory, President Barack Obama won
­re-election in 2012 with 332 electoral votes versus G
­ overnor
Romney’s 206. This was in stark contrast to the Romney victory confidently predicted by Republican pundits just days
before the election. Perhaps this was a result of a Gallup
poll weeks before the election that showed Romney with
a lead in the popular vote. However, American presidents
are not elected by popular vote. They are elected by the
Electoral College, which vote on a state by state basis. Even
though several past Presidents have won the election while
losing the popular vote, this was not the case in 2012.
A statistician named Nate Silver predicted the outcome of the election exactly. He based his forecasts on
careful analysis of state by state (and even county by
county) polls. His methodology had previously proven
very ­successful in baseball. He also accurately predicted
the outcome of the 2008 presidential election in which he
was correct in 49 out of 50 states. But in 2012 he was correct in all fifty states, predicting not only an Obama victory with probability exceeding 90 percent, but predicting
the number of electoral votes for each candidate exactly.
Silver’s striking successes points to the importance in
using sound analytics for forecasting. His is not an isolated case, however. Reports abound about such methods being successfully applied in many other contexts.
(Note that a group from Princeton University, applying
similar methods, was also able to accurately predict the
outcome of the 2012 election.)
Year 1DemandYear 2DemandYear 3Demand
1
2
3
4
12
25
76
52
1
2
3
4
16
32
71
62
1
2
3
4
14
45
84
47
a. Determine the seasonal factors for each quarter.
b. Based on the result of part (a), determine the deseasonalized demand s­ eries.
c. Predict the demand for each quarter of Year 4 for the deseasonalized series from
a six-quarter moving average.
d. Using the results from parts (a) and (c), predict the demand for the shoes for
each quarter of Year 4.
Winters’s Method for Seasonal Problems
The moving-average method just described can be used to predict a seasonal series
with or without a trend. However, as new data become available, the method requires
that all seasonal factors be recalculated from scratch. Winters’s method is a type of
triple exponential smoothing, and this has the important advantage of being easy to
­update as new data become available.
We assume a model of the form
Dt 5 (m 1 Gt)ct 1 et .
Interpret m as the base signal or intercept at time t 5 0 excluding seasonality, Gt as the
trend or slope component, ct as the multiplicative seasonal component in period t, and
finally et as the error term. Because the seasonal factor multiplies both the base level
and the trend term, we are assuming that the underlying series has a form similar to
that pictured in Figure 2–10.
Again assume that the length of the season is exactly N periods and that the ­seasonal
factors are the same each season and have the property that S ct 5 N. Three exponential
nah77856_ch02.indd 85
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 85
2.9
86 Chapter Two Forecasting
Methods for Seasonal Series 85
FIGURE 2–10
Observations
Seasonal series with
increasing trend
Time
smoothing equations are used each period to update estimates of deseasonalized series,
the seasonal factors, and the trend. These equations may have different smoothing constants, which we will label �, �, and �.
1. The series. The current level of the deseasonalized series, St, is given by
St � �(Dtct�N) � (1 � �)(St�1 � Gt�1).
Notice what this equation does. By dividing by the appropriate seasonal factor, we are
deseasonalizing the newest demand observation. This is then averaged with the current
forecast for the deseasonalized series, as in Holt’s method.
2. The trend. The trend is updated in a fashion similar to Holt’s method.
Gt � �[St � St�1] � (1 � �)Gt�1.
3. The seasonal factors.
ct � �(DtSt) � (1 � �)ct�N.
The ratio of the most recent demand observation over the current estimate of the
deseasonalized demand gives the current estimate of the seasonal factor. This is then
averaged with the previous best estimate of the seasonal factor, ct�N. Each time that
a seasonal factor is updated, it is necessary to norm the most recent N factors to add to N.
Finally, the forecast made in period t for any future period t � � is given by
Ft, t�� � (St � �Gt)ct���N.
Note that this forecasting equation assumes that t � N. If N � � � 2N, the appropriate
seasonal factor would be ct���2N; if 2N � � � 3N, the appropriate seasonal factor
would be ct���3N; and so on.
nah77856_ch02.indd 86
17/12/14 6:47 PM
nah77856_ch02.qxd
86
12/31/07
4:21 PM
Page 86
Chapter Two Forecasting
2.9 Methods for Seasonal Series 87
Initialization Procedure
In order to get the method started, we need to obtain initial estimates for the series, the
slope, and the seasonal factors. Winters suggests that a minimum of two seasons of data
be available for initialization. Let us assume that exactly two seasons of data are available; that is, 2N data points. Suppose that the current period is t � 0, so that the past
observations are labeled D�2N�1, D�2N�2, . . . , D0.
1. Calculate the sample means for the two separate seasons of data.
V1 �
�N
1
a Dj
N j� �2N�1
V2 �
1
N
a
0
Dj
j� �N�1
2. Define G0 � (V2 � V1)N as the initial slope estimate. If m � 2 seasons of
data are available for initialization, then compute V1, . . . , Vm as in step 1 and define
G0 � (Vm � V1)[(m � 1)N ]. If we locate V1 at the center of the first season of data [at
period (�3N � 1)2] and V2 at the center of the second season of data [at period
(�N � 1)2], then G0 is simply the slope of the line connecting V1 and V2 (refer to
Figure 2–11).
3. Set S0 � V2 � G0[(N � 1)2]. This estimates the value of the series at time
t � 0. Note that S0 is the value assumed by the line connecting V1 and V2 at t � 0
(see Figure 2–11).
4. a. The initial seasonal factors are computed for each period in which data are
available and then averaged to obtain one set of seasonal factors. The initial seasonal
factors are obtained by dividing each of the initial observations by the corresponding
FIGURE 2–11
20
19
18
17
16
15
14
13
12
11
10
9
V2 8
7
6
V1 5
4
3
2
1
0
Demand
Initialization for
Winters’s method
nah77856_ch02.indd 87
Season 1
Season 2
Slope = G0
Time
N
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 87
2.9
88 Chapter Two Forecasting
Methods for Seasonal Series 87
point along the line connecting V1 and V2. This can be done graphically or by using the
following formula:
ct �
Dt
Vi � [(N � 1)>2 � j]G0
for �2N � 1 … t … 0,
where i � 1 for the first season, i � 2 for the second season, and j is the period of the
season. That is, j � 1 for t � �2 N � 1 and t � �N � 1; j � 2 for t � �2N � 2 and
t � �N � 2; and so on.
b. Average the seasonal factors. Assuming exactly two seasons of initial data, we
obtain
c �2N�1 � c �N�1
c �N � c0
, . . . , c0 �
.
c�N�1 �
2
2
c. Normalize the seasonal factors.
cj �
cj
J a ci K
�N�1
#N
for �N � 1 … j … 0.
i�0
This initialization procedure just discussed is the one suggested by Winters. It is not
the only means of initializing the system. Seasonal factors could be determined by the
method of moving averages discussed in the first part of this section. Another alternative would be to fit a linear regression to the baseline data and use the resulting slope
and intercept values, as was done in Holt’s method, to obtain S0 and G0. The seasonal
factors would be obtained by dividing each demand observation in the baseline period
by the corresponding value on the regression line, averaging like periods, and norming. The actual values of the initial estimates of the intercept, the slope, and the seasonal factors will be similar no matter which initialization scheme is employed.
Example 2.8
Assume that the initial data set is the same as that of Example 2.7, in which centered moving
averages were used to find the seasonal factors. Recall that we have two seasons of data: 10,
20, 26, 17, 12, 23, 30, 22. Then
V1 � (10 � 20 � 26 � 17)4 � 18.25,
V2 � (12 � 23 � 30 � 22)4 � 21.75,
G0 � (21.75 � 18.25)4 � 0.875,
S0 � 21.75 � (0.875)(1.5) � 23.06.
The initial seasonal factors are computed as follows:
10
� 0.5904,
18.25 � (52 � 1)(0.875)
20
�
� 1.123.
18.25 � (52 � 2)(0.875)
c�7 �
c�6
The other factors are computed in a similar fashion. They are
nah77856_ch02.indd 88
c�5 � 1.391,
c�4 � 0.869,
c�3 � 0.5872,
c�2 � 1.079,
c�1 � 1.352,
c0 � 0.9539.
17/12/14 6:47 PM
nah77856_ch02.qxd
88
12/31/07
4:21 PM
Page 88
Chapter Two Forecasting
2.9 Methods for Seasonal Series 89
We then average c�7 and c�3, c�6 and c�2, and so on, to obtain the four seasonal factors:
c�3 � 0.5888,
c�2 � 1.1010,
c�1 � 1.3720,
c0 � 0.9115.
Finally, norming the factors to ensure that the sum is 4 results in
c�3 � 0.5900,
c�2 � 1.1100,
c�1 � 1.3800,
c0 � 0.9200.
Notice how closely these factors agree with those obtained from the moving-average method.
Suppose that we wish to forecast the following year’s demand at time t � 0. The forecasting equation is
Ft,t� � (St � Gt)ct� �N,
which results in
F0,1 � (S0 � G0)c�3 � (23.06 � 0.875)(0.59) � 14.12,
F0,2 � (S0 � 2G0)c�2 � [23.06 � (2)(0.875)](1.11) � 27.54,
F0,3 � 35.44,
F0,4 � 24.38.
Now, suppose that at time t � 1 we observe a demand of D1 � 16. We now need to update
our equations. Assume that � .2, � .1, and � .1. Then
S1 � (D1c�3) � (1 � )(S0 � G0)
� (0.2)(16.59) � (0.8)(23.06 � 0.875) � 24.57,
G1 � (S1 � S0) � (1 � )G0
� (0.1)(24.57 � 23.06) � (0.9)(0.875) � 0.9385,
c1 � (D1S1) � (1 � )c�3
� (0.1)(1624.57) � (0.9)(0.59) � 0.5961.
At this point, we would renorm c�2, c�1, c0, and the new value of c1 to add to 4. Because the
sum is 4.0061, it is close enough (the norming would result in rounding c1 down to .59 once
again).
Forecasting from period 1, we obtain
F1,2 � (S1 � G1)c�2 � (24.57 � 0.9385)(1.11) � 28.3144,
F1,3 � (S1 � 2G1)c�1 � [24.57 � (2)(0.9385)](1.38) � 36.4969,
and so on.
Now suppose that we have observed one full year of demand, given by D1 � 16, D2 � 33,
D3 � 34, and D4 � 26. Each time a new observation becomes available, the intercept, slope,
and most current seasonal factor estimates are updated. One obtains
S2 � 26.35,
S3 � 26.83,
S4 � 27.89,
G2 � 1.0227,
G3 � 0.9678,
G4 � 0.9770,
c2 � 1.124,
c3 � 1.369,
c4 � 0.9212.
As c1, c2, c3, and c4 sum to 4.01, normalization is not necessary. Suppose that we were interested in the forecast made in period 4 for period 10. The forecasting equation is
Ft,t� � (St � Gt)ct��2N,
which results in
F4,10 � (S4 � 6G4)c2 � [27.89 � 6(0.9770)](1.124) � 37.94.
nah77856_ch02.indd 89
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 89
2.10
90 Chapter Two Forecasting
Box-Jenkins Models 89
An important consideration is the choice of the smoothing constants , , and to be used
in Winters’s method. The issues here are the same as those discussed for simple exponential
smoothing and Holt’s method. Large values of the smoothing constants will result in more responsive but less stable forecasts. One method for setting , , and is to experiment with various values of the parameters that retrospectively give the best fit of previous forecasts to the
observed history of the series. Because one must test many combinations of the three constants, the calculations are tedious. Furthermore, there is no guarantee that the best values of
the smoothing constants based on past data will be the best values for future forecasts. The
most conservative approach is to guarantee stable forecasts by choosing the smoothing constants to be between 0.1 and 0.2.
More Problems for Section 2.9
35. Consider the data for Problem 34.
a. Using the data from Years 2 and 3, determine initial values of the intercept,
slope, and seasonal factors for Winters’s method.
b. Assume that the observed demand for the first quarter of Year 4 was 18. Using
� .2, � .15, and � .10, update the estimates of the series, the slope, and
the seasonal factors.
c. What are the forecasts made at the end of the first quarter of Year 4 for the
remaining three quarters of Year 4?
36. Suppose the observed quarterly demand for Year 4 was 18, 51, 86, 66. Compare the
accuracy of the forecasts obtained for the last three quarters of Year 4 in Problems
34(d ) and 35(c) by computing both the MAD and the MSE.
37. Determine updated estimates of the slope, the intercept, and the seasonal factors
for the end of Year 4 based on the observations given in Problem 36. Using these
updated estimates, determine the forecasts that Winters’s method gives for all of
Year 6 made at the end of Year 4. Use the values of the smoothing constants given
in Problem 35(b).
2.10
BOX-JENKINS MODELS
The forecasting models introduced in this section are significantly more sophisticated
than those previously discussed in this chapter. The goal is to present the basic concepts of Box-Jenkins analysis so that the reader can appreciate the power of these
methods. However, an in-depth coverage is beyond the scope of this book. The
methods are named for two well-known statisticians, George E. Box and Gwilym M.
Jenkins formerly from the University of Wisconsin and University of Lancaster, respectively. The approach they developed is based on exploiting the autocorrelation
structure of a time series. While Box-Jenkins methods are based on statistical relationships in a time series, much of the basic theory goes back to the famous book by
Norbert Wiener (1949), and before.
Box-Jenkins models are also known as ARIMA models. ARIMA is an acronym for
autoregressive integrated moving average. The autocorrelation function plays a central
role in the development of these models, and is the feature that distinguishes ARIMA
models from the other methods discussed in this chapter. As we have assumed throughout this chapter, denote the time series of interest as D1, D2, . . . . We will assume
initially that the series is stationary. That is, E1Di) � and Var(Di) � 2 for all
i � 1, 2, . . . . Practically speaking, stationarity means that there is no growth or
nah77856_ch02.indd 90
17/12/14 6:47 PM
nah77856_ch02.qxd
90
12/31/07
4:21 PM
Page 90
Chapter Two Forecasting
2.10 Box-Jenkins Models 91
decline in the series, and variation remains relatively constant. It is important to note
that stationarity does not imply independence. Hence, it is possible that values of Di
and Dj are dependent random variables when i Z j even though their marginal density
functions are the same. It is this dependence we wish to exploit. (Note: A more precise way to characterize stationarity is that the joint distribution of Dt, Dt1, . . . , Dtk
is the same as the joint distribution of Dtm, Dtm1, . . . , Dtmk at any time t and
pair of positive integers m and k.)
The assumption of stationarity implies that the marginal distributions of any two
observations separated by the same time interval are the same. That is, Dt and Dt1
have the same joint distribution as Dtm and Dtm1 for any m Ú 1. This implies that
the covariance of Dt and Dt1 is exactly the same as the covariance of Dtm and
Dtm1. Hence, the covariance of any two observations depends only on the number of
periods separating them. In this context, the covariance is also known as the autocovariance, since we are comparing two values of the same series separated by a fixed lag.
Let Cov (Dtm, Dtmk) be the covariance of Dtm and Dtmk given by
Cov(Dtm, Dtmk) E(DtmDtmk) E(Dtm)E(Dtmk)
for any integer k Ú 1.
The correlation coefficient of these two random variables is given by
Cov(Dtm, Dtmk)
.
2Var(Dtm)2Var(Dtmk)
This is often referred to as the autocorrelation coefficient of lag k, since it refers
to the correlation between all values of the series separated by k periods. These
autocorrelation coefficients are typically computed for several values of k. It is these
autocorrelation coefficients that will play the key role in building ARIMA models.
The autocorrelation coefficients are estimated from a history of the series. In order
to guarantee reliable estimators, Box and Jenkins (1970) suggest that one have at least
72 data points of past history of the series. Hence, these models are only meaningful
when one has a substantial and reliable history of the series being studied.
k Estimating the Autocorrelation Function
Let D1, D2, . . . , Dn be a history of observations of a time series. The autocorrelation
coefficient of lag k is estimated from the following formula:
a (Dt D)(Dtk D)
n
rk tk1
2
a (Dt D)
n
,
t1
where D is the sample mean (that is, the average) of the observed values of the series.
Refer to the rk as sample autocorrelation coefficients. This calculation is typically done
for 10 or 15 values of k. For most of the time series discussed earlier in the chapter, one
identifies the appropriate patterns by just looking at a graph of the data. This is not the
case here, however.
Example 2.9
nah77856_ch02.indd 91
If observations are completely random (i.e., form a white noise process), then we would expect
that there would be no significant autocorrelations among the observed values of the series. To
test this, we generated a time series using the random number generator built into Excel. This
series appears in Table 2–1. Each value is 100 times the RAND function. The reader can check
that the sample autocorrelations for lags of 1 to 10 periods for these 36 observations are
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 91
2.10
92 Chapter Two Forecasting
TABLE 2–1
Time Series with
36 Values Generated
by a Random
Number Generator
(White Noise
Series)
Period
1
2
3
4
5
6
7
8
9
10
11
12
Box-Jenkins Models 91
Value
Period
Value
Period
Value
42
93
17
5
38
2
67
66
11
65
88
91
13
14
15
16
17
18
19
20
21
22
23
24
47
52
28
58
41
47
48
50
81
93
45
24
25
26
27
28
29
30
31
32
33
34
35
36
88
73
60
56
49
51
59
80
40
60
20
35
r1 r2 r3 r4 r5 r6 r7 r8 r9 r10 0.098
0.118
0.018
0.080
0.0752
0.006
0.270
0.207
0.117
0.136
If we had an infinite number of observations and a perfect white noise series, we would expect
that all of the autocorrelations would be zero. However, since we only have a finite series, there
will be statistical variation resulting in nonzero values of the autocorrelations. The question is
whether these values are significantly different from zero. (The data from Table 2–1 appear in
Figure 2–12 and the autocorrelations in Figure 2–13.)
FIGURE 2–12
Plot of data in
Table 2–1
100
90
80
70
Value
60
50
40
30
20
10
0
0
10
20
30
40
Time
nah77856_ch02.indd 92
17/12/14 6:47 PM
nah77856_ch02.qxd
92
12/31/07
4:21 PM
Page 92
Chapter Two Forecasting
2.10 Box-Jenkins Models 93
FIGURE 2–13
Plot of autocorrelations
of series in Figure 2–12
0.2
0.15
Autocorrelation
0.1
0.05
0
–0.05
0
2
4
6
8
10
12
–0.1
–0.15
–0.2
–0.25
–0.3
Lag
Several statistical tests have been proposed to answer this question. One is the Box-Pierce Q
statistic, computed from the formula
Q n a rk2,
h
k 1
where h is the maximum length of the lag being considered, and n is the number of observations in the series. Under the null hypothesis that the series is white noise, the Q statistic has
the chi-square distribution with (h m) degrees of freedom, where m is the number of
parameters in the model that has been fitted to the data. This test can be applied to any set of
data not necessarily fitted to a specific model by setting m 0.
Applying the formula for the Q statistic to the preceding autocorrelations, we obtain a value
of Q 6.62. Comparing this to the critical values of the chi-square statistic in Table 2–2 with
10 degrees of freedom (df), we see that this value is substantially smaller than any value in the
table (for example, the value for a right tail probability of .1 at 10 degrees of freedom is 15.99).
Hence, we could not reject the null hypothesis that the data form a white noise process, and
we conclude that the autocorrelations are not significant.
TABLE 2–2
Partial Table of
Critical Values of
the Chi-Square
Statistic
nah77856_ch02.indd 93
Tail Value Probability
df
0.1
0.05
0.01
1
2
3
4
5
6
7
8
9
10
2.70
4.60
6.25
7.77
9.23
10.64
12.01
13.36
14.68
15.99
3.84
5.99
7.81
9.48
11.07
12.59
14.06
15.50
16.91
18.30
6.63
9.21
11.34
13.27
15.08
16.81
18.47
20.09
21.66
23.20
17/12/14 6:47 PM
nah77856_ch02_052-123.qxd
2/17/10
5:13 PM
Page 93
2.10
94 Chapter Two Forecasting
Box-Jenkins Models 93
Of course, the real interest is in those cases where the autocorrelations are significant. The
basic idea behind the method is to compare the graph of the autocorrelation function (known
as the correlogram) to those of known processes to identify the appropriate model. (Note: In
addition to considering autocorrelations, most texts on Box-Jenkins analysis also recommend
examining the partial autocorrelations. We will not discuss partial autocorrelations here, but
the reader should be aware that these also provide information about the underlying structure
of the process.)
The Autoregressive Process
The autoregressive model is
Dt a0 a1Dt1 a2Dt2 Á apDtp Pt,
where a0, a1, . . . , ap are the linear regression coefficients and t is the error term
(generally assumed to be normal with mean 0 and variance 2 as earlier in the chapter). The reader familiar with linear regression will recognize this equation as being
very similar to the standard regression equation with Dt playing the role of the
dependent variable and Dt1, Dt2, . . . , Dtp playing the role of the independent variables. Hence, the autoregressive model regresses the value of the series at time t on the
values of the series at times t 1, t 2, . . . , t p. Note, however, that there is a fundamental difference between an autoregressive model and a simple linear regression,
since in this case it is likely that the variables are correlated. We will use the notation
AR( p) to represent this model.
Consider a basic AR(1) model,
Dt a0 a1Dt1 t.
In order for the process to be stable, we require |a1| 6 1. If a1 7 0, it means that successive values of the series are positively correlated—that is, large values tend to be
followed by large values, and small values tend to be followed by small values. This
means that the series will be relatively smooth. If a1 6 0, then the opposite is true, so
the series will appear much spikier. The difference is illustrated in a realization of two
AR(1) processes in Figure 2–14. (Figure 2–14 was generated in Excel using the builtin RAND function and the normal variate generator given in Problem 5.34 with
a0 10 and 30.)
Of course, it is unlikely one can recognize an AR(1) process by simply examining a
graph of the raw data. Rather, one would examine the autocorrelation function. It is
easy to show that the autocorrelation function for an AR(1) process (see Nelson, 1973,
page 39, for example) is
j a1 j.
The autocorrelation functions for the two cases illustrated in Figure 2–14 are given in
Figure 2–15. If the sample autocorrelation function of a series has a pattern resembling
one of those in Figure 2–15, it would suggest that an AR(1) process is appropriate. The
theoretical autocorrelation functions for higher-order AR processes can be more complex. [In the case of AR(2), the patterns are either similar to one of the two pictured in
Figure 2–15 or follow a damped sine wave.] In practice, one would rarely include more
than one or two AR terms in the model. Determining the autocorrelation structure for
higher-order AR processes is not difficult. One must solve a series of linear equations
known as the Yule-Walker equations. We will not elaborate further here, but refer the
interested reader to Box and Jenkins (1970).
nah77856_ch02.indd 94
17/12/14 6:47 PM
nah77856_ch02.qxd
94
12/31/07
4:21 PM
Page 94
Chapter Two Forecasting
2.10 Box-Jenkins Models 95
FIGURE 2–14
Realizations of an
AR(1) process with
a1 0.9 and a1 0.9,
respectively.
120
100
80
60
40
20
0
–20
0
10
20
30
40
50
60
40
50
60
–40
–60
a1 = 0.9
200
150
100
50
0
0
10
20
30
–50
–100
–150
a1 = –0.9
The Moving-Average Process
The moving-average process provides another means of describing a stationary stochastic process used to model time series. The term moving average as used here
should not be confused with the moving average discussed earlier in this chapter in
Section 2.7. In that case, the moving average was an average of past values of the
series. In this case, the moving average is a weighted average of past forecast errors.
nah77856_ch02.indd 95
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 95
2.10
96 Chapter Two Forecasting
Box-Jenkins Models 95
FIGURE 2–15
Theoretical autocorrelation functions for AR(1) processes pictured in Figure 2–14.
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20
1
0.8
0.6
0.4
0.2
0
–0.2
–0.4
–0.6
–0.8
–1
The general moving-average process has the form
Dt b0 b1t1 b2t2 . . . bqtq t.
[The weights b1, b2, . . . , bq are shown with negative signs by convention.] We will
denote this model as MA(q). The intuition behind the moving-average process is not
nah77856_ch02.indd 96
17/12/14 6:47 PM
nah77856_ch02.qxd
96
12/31/07
4:21 PM
Page 96
Chapter Two Forecasting
2.10 Box-Jenkins Models 97
as straightforward as the autoregressive process, but the two are related. Consider the
first-order AR(1) process, Dt a0 a1Dt1 t. By back-substituting for Dt1,
Dt2, . . . , we see that the AR(1) process can also be written as
Dt a0 a ai1 a1t1 a21t2 . . . t,
q
i0
which is easily recognized as an MA(q) process.
The reader will better understand the power of MA processes when we examine the
autocorrelation function. Consider first the simplest MA process, namely, an MA(1)
process. In this case,
Dt b0 b1t1 t.
The autocorrelation structure for this case is very simple:
b1
1 ,
1 b21
3 Á 0.
2
Hence, an MA(1) process has only one significant autocorrelation at lag 1. In general,
MA(1) processes tend to have spiky patterns independent of the sign of b1, because
successive errors are assumed to be uncorrelated. Figure 2–16 shows realizations of an
MA(1) process with b1 equal to 0.9 and 0.9, respectively. Finding the autocorrelation
structure for higher-order MA processes is a challenging mathematical problem, requiring the solution of a collection of nonlinear equations. Again, we refer the interested reader to Box and Jenkins (1970). The main characteristic identifying an MA(q)
process is that only the first q autocorrelations are nonzero.
Mixtures: ARMA Models
The real power in Box-Jenkins methodology comes in being able to mix both AR and
MA terms. Any model that contains one or more AR terms and one or more MA terms
is known as an ARMA model, for autoregressive moving average. An ARMA(p, q)
model contains p autoregressive terms and q moving-average terms. For example, one
would write an ARMA(1,1) model as
Dt c a1Dt1 b1t1 t.
The ARMA(1,1) model is quite powerful, and can describe many real processes accurately. It requires identification of the two parameters a1 and b1. The ARMA(1, 1) process
is equivalent to an MA(q) process and also equivalent to an AR(q) process, thus showing the power that can be achieved with a parsimonious model. The autocorrelation
function of the ARMA(1,1) process has characteristics of both the MA(1) and AR(1)
processes. The autocorrelation at lag 1 is determined primarily by the MA(1) term, while
autocorrelations at lags greater than 1 are determined by the AR(1) term.
ARIMA Models
Thus far we have assumed that the underlying stochastic process generating the time
series is stationary. However, in many real problems, patterns such as trend or seasonality are present, which imply nonstationarity. The Box-Jenkins approach would be of
limited utility if it were unable to address such situations. Fortunately, there is a simple
technique for converting many nonstationary processes into stationary processes.
nah77856_ch02.indd 97
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 97
2.10
98 Chapter Two Forecasting
Box-Jenkins Models 97
FIGURE 2–16
Realizations of an
MA(1) process with
b1 0.9 and
b1 0.9, respectively
120
100
80
60
40
20
0
–20
0
10
20
30
40
50
60
40
50
60
–40
–60
–80
–100
b1 = –0.9
200
150
100
50
0
0
10
20
30
–50
–100
–150
b1 = 0.9
Consider first a process with a linear trend, such as the one pictured in Figure 2–7.
Simple methods for dealing with linear trends were discussed in Section 2.8 of this
chapter. How can a process with a linear trend be converted to one with no trend? The
answer turns out to be surprisingly simple. Suppose our original process Dt has a linear
trend. Consider the new process Ut given by
Ut Dt Dt1.
nah77856_ch02.indd 98
17/12/14 6:47 PM
nah77856_ch02.qxd
98
12/31/07
4:21 PM
Page 98
Chapter Two Forecasting
2.10 Box-Jenkins Models 99
The process Ut tracks the slope of the original process. If the original process had a linear trend, then the slope should be relatively constant, implying that Ut would be stationary. In the same way, if the original process increased or decreased according to a
quadratic function, differencing the first difference process (forming a second difference) will result in a stationary process. Differencing is the discrete analogue of a derivative. Going from the process Ut back to the original process Dt requires summing
values of Ut , which is the discrete analogue of integration. For that reason, when differencing is introduced, we use the term integration to describe it. An ARMA model
based on data derived from differencing is denoted ARIMA, which stands for autoregressive integrated moving average.
A common notation is Ut §Dt. If two levels of differencing were required to
achieve stationarity, then one would need a double summation to retrieve the original
series. In the case of two levels of differencing,
§2Dt Dt Dt1 (Dt1 Dt2) Dt 2Dt1 Dt2.
Differencing can also be used to remove seasonality from a time series. Suppose that a
seasonal pattern repeats every 12 months. Then defining
Ut §12Dt Dt Dt12
would result in a process with no seasonality.
An ARIMA process has three constants associated with it: p for the number of
autoregressive terms, d for the order of differencing, and q for the number of movingaverage terms. The general ARIMA process would be denoted ARIMA(p, d, q).
Thus, for example, ARMA(1,1) can also be denoted ARIMA(1,0,1). While these
parameters can be any nonnegative integers, it is very rare that any of the values of
p, d, or q would exceed 2. Thus, virtually all the ARIMA models one finds in practice correspond to values of 0, 1, or 2 for the parameters p, d, and q. While this
might seem limiting, these few cases cover an enormous range of practical forecasting scenarios.
It is important to note that observations are lost when differencing. For example, if
one uses a single level of differencing (Ut Dt Dt1), the first difference process,
Ut, will have one less observation than the original series. Similarly, each level of differencing will reduce the sample size by 1. Seasonal differencing effectively reduces
the data set by the length of the season. Another way to represent the differencing
operation is via the backshift operator. That is,
BDt Dt1,
which means that one would represent the first difference process as
Ut Dt Dt1 Dt BDt (1 B)Dt.
The backshift operator can also be used to simplify the notation for AR, MA, and
ARMA models. Consider the simple AR(1) process given by Dt a0 a1Dt1 t.
Writing this in the form Dt a1Dt1 a0 t reduces the process to the alternate
notation (1 a1B)Dt a0 t. Similarly, the reader should check that the MA(1)
model using the backshift operator is Dt b0 (1 b1B)et.
Using ARIMA Models for Forecasting
Given an ARIMA model, how does one go about using it to provide forecasts of future
values of the series? The approach is similar to that discussed earlier in this chapter
for the simpler methods. For example, a forecast based on a simple AR(p) model is
nah77856_ch02.indd 99
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 99
2.10
100 Chapter Two Forecasting
Box-Jenkins Models 99
a weighted average of past p observations of the series. A forecast based on an
MA(q) model is a weighted average of the past q forecast errors. And finally, one must
take into account the level of differencing and any transformations made on the
original data.
As an example, consider an ARIMA (1, 1, 1) model (that is, a model having one
level of differencing, one moving-average term, and one autoregressive term). This can
be represented using the backshift notation as
(1 B)(1 a1B)Dt c (1 b1B)t,
or as
(1 a1B)§Dt c (1 b1B)t.
Writing out the model without backshift notation we have
Dt c (1 a1)Dt1 a1Dt2 t b1t1.
Let us suppose this model has been fitted to a time series with the result that c 15,
a1 0.24, and b1 0.70. Suppose that the last five values of the time series used to fit
this model were 31.68, 29.10, 43.15, 56.74, and 62.44 based on a total of 76 observations. The first period in which one can forecast is period 77. The one-step-ahead forecast for period 77 made at the end of period 76 is
DN 77 15 (1 0.24)D76 0.24D75 0.7076.
DN 77 is the conditional expected value of D77 having observed the demand in periods 1,
2, . . . , 76. Since this is the first forecast made for this series, there is no observed previous
value of the error, and the final term drops out. Hence, the one-step-ahead forecast made
in period 76 for the demand in period 77 is 15 (1.24)(62.44) (0.24)(56.74) =
78.81. Now, suppose we observed a value of 70 for the series in period 77. That means
that the forecast error observed in period 77 is 77 78.81 70 8.81. The onestep-ahead forecast for period 78 made in period 77 would be 15 (1.24)(70) 0.24(62.44) 0.70(8.81) 86.25.
When using an ARIMA model for multiple-step-ahead forecasts, the operative rule
is to use the forecasts for the unobserved demand and use zero for the unobserved
errors. Thus in the preceding example, a two-step-ahead forecast made at the end of
period 76 for demand in period 78 would be based on the assumption that the observed
demand in period 77 was the one-step-ahead forecast, 86.25. The observed forecast
error for period 77 would be assumed to be zero.
Summary of the Steps Required for Building ARIMA Models
There are four major steps required for building Box-Jenkins forecasting models.
1. Data transformations. The Box-Jenkins methodology is predicated on starting
with a stationary time series. To be certain that the series is indeed stationary, several preliminary steps might be required. We know that differencing eliminates trend and seasonality. However, if the mean of the series is relatively fixed, it still may be the case that
the variance is not constant, thus possibly requiring a transformation of the data (for
example, stock market data are often transformed by the logarithm).
2. Model identification. This step refers to determining exactly which ARIMA
model seems to be most appropriate. Proper model identification is both art and science. It is difficult, if not impossible, to identify the appropriate model by only examining the series itself. It is far more effective to study the sample autocorrelations and
nah77856_ch02.indd 100
17/12/14 6:47 PM
nah77856_ch02.qxd
100
12/31/07
4:21 PM
Page 100
Chapter Two Forecasting
2.10 Box-Jenkins Models 101
partial autocorrelations to discern patterns that match those of known processes. In
some cases, the autocorrelation structure will point to a simple AR or MA process, but
it is more common that some mixture of these terms would be required to get the best
fit. However, one must not add terms willy-nilly. The operative concept is parsimony—
that is, use the most economical model that adequately describes the data.
3. Parameter estimation. Once the appropriate model has been identified, the optimal
values of the model parameters (i.e., a0, a1, . . . , ap and b0, b1, . . . , bq) must be determined. Typically, this is done via either least squares fitting methods or the method of
maximum likelihood. In either case, this step is done by a computer program.
4. Forecasting. Once the model has been identified and the optimal parameter values
determined, the model provides forecasts of future values of the series. Box-Jenkins
models are most effective in providing one-step-ahead forecasts, but can also provide
multiple-step-ahead forecasts as well.
5. Evaluation. The pattern of residuals (forecast errors) can provide useful information regarding the quality of the model. The residuals should form a white noise
(i.e., random) process with zero mean. Residuals should be normally distributed as
well. When there are patterns in the residuals, it suggests that the forecasting model
can be improved.
Case Study. Using Box-Jenkins Methodology to Predict
Monthly International Airline Passenger Totals
This study is based on data that appeared originally in Brown (1962), but was analyzed using
ARIMA methods in Box and Jenkins (1970). It illustrates the basic steps in transforming
data and building ARIMA models. The data represent the monthly international airline sales
from the period January 1949 to December 1960. The raw data appear in Table 2–3 and are
pictured in Figure 2–17. From the figure, it is clear that there are several nonstationarities
in this data. First, there is clearly an increasing linear trend. Second, there is seasonality,
with a pattern repeating yearly. Third, there is increasing variance over time. In cases where
the mean and variance increase at a comparable rate (which would occur if the series is
TABLE 2–3
International Airline
Passengers: Monthly
Totals (Thousands of
Passengers), January
1949–December
1960*
Source: From Box and Jenkins
(1970), p. 531.
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
Jan. Feb.
Mar. Apr.
May June
July
Aug. Sept. Oct. Nov.
Dec.
112
115
145
171
196
204
242
284
315
340
360
417
132
141
178
193
236
235
267
317
356
362
406
419
121
125
172
183
229
234
270
318
355
363
420
472
148
170
199
230
264
302
364
413
465
491
548
622
148
170
199
242
272
293
347
405
467
550
559
606
118
140
166
194
201
229
278
306
336
337
405
432
118
126
150
180
196
188
233
277
301
318
342
391
129
135
163
181
235
227
269
313
348
348
396
461
135
149
178
218
243
264
315
374
422
435
472
535
136
158
184
209
237
259
312
355
404
404
463
508
119
133
162
191
211
229
274
306
347
359
407
461
104
114
146
172
180
203
237
271
305
310
362
390
*
144 observations.
nah77856_ch02.indd 101
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 101
2.10
102 Chapter Two Forecasting
FIGURE 2–17
International airline
passengers (thousands)
Box-Jenkins Models 101
700
600
500
400
300
200
100
FIGURE 2–18
Natural log of
international airline
passengers
57
Ja
n58
Ja
n59
Ja
n60
Ja
n-
56
55
Ja
n-
Ja
n-
54
53
Ja
n-
Ja
n-
52
51
Ja
n-
50
Ja
n-
Ja
n-
Ja
n-
49
0
6.500
6.300
6.100
5.900
5.700
5.500
5.300
5.100
4.900
4.700
60
59
Ja
n-
58
Ja
n-
Ja
n-
57
Ja
n-
56
Ja
n-
55
Ja
n-
54
Ja
n-
53
Ja
n-
52
Ja
n-
51
Ja
n-
50
Ja
n-
Ja
n-
49
4.500
increasing by a fixed percentage), a logarithmic transformation will usually eliminate the
nonstationarity due to increasing variance. (Changing variance is known as heteroscedasticity, and constant variance is known as homoscedasticity.) Applying a natural log transformation to the data yields a homoscedastic series as shown in Figure 2–18.
Next, we need to apply two levels of differencing to eliminate both trend and seasonality. The trend is eliminated by applying a single level of differencing and the seasonality by
applying 12 periods of differencing. After these three transformations are applied to the
original data, the resulting data appear in Figure 2–19. Note now the transformed data
nah77856_ch02.indd 102
17/12/14 6:47 PM
nah77856_ch02.qxd
102
12/31/07
4:21 PM
Page 102
Chapter Two Forecasting
2.10 Box-Jenkins Models 103
FIGURE 2–19
Natural log of airline
data with single and
seasonal differencing
0.2
0.15
0.1
Value
0.05
0
0
20
40
60
80
100
120
140
–0.05
–0.1
–0.15
–0.2
Time period
Series1
appear to form a random white noise process centered at zero showing neither trend nor
seasonality.
It is to this set of data that we wish to fit an ARMA model. To do so, we determine the
sample autocorrelations. While this can be accomplished with many of the software programs available for forecasting (including general statistical packages, such as SAS), we
have done the calculations directly in Excel using the formulas for sample autocorrelations
given earlier. The autocorrelations for lags of 1 to 12 periods for the series pictured in Figure 2–19 appear in Table 2–4.
Although not explicitly discussed in this section, when both seasonal differencing and
period-to-period differencing are applied simultaneously, one must determine an ARMA
model for each level of differencing. That is, one would like to find an ARMA model
corresponding to the single level of differencing and to the seasonal level of differencing.
Thus, when examining the autocorrelation function, we look for patterns at lags of 1 period and patterns at lags of 12 periods. From Table 2–4 it is clear that there are significant
TABLE 2–4
Autocorrelations for
the Transformed
Airline Data Pictured
in Figure 2–19 (after
taking logarithms
and two levels of
differencing)
nah77856_ch02.indd 103
Lag
1
2
3
4
5
6
7
8
9
10
11
12
Autocorrelation
Lag
Autocorrelation
Lag
Autocorrelation
0.34
0.11
0.20
0.02
0.06
0.03
0.06
0.00
0.18
0.08
0.06
0.39
13
14
15
16
17
18
19
20
21
22
23
24
0.15
0.06
0.15
0.14
0.07
0.02
0.01
0.12
0.04
0.09
0.22
0.02
25
26
27
28
29
30
31
32
33
34
35
36
0.10
0.05
0.03
0.05
0.02
0.05
0.05
0.20
0.12
0.08
0.15
0.01
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 103
2.10
104 Chapter Two Forecasting
Box-Jenkins Models 103
autocorrelations at lags of exactly 1 and 12 periods. This suggests that MA(1) models are
appropriate for both differencing levels.
Note that if we let zt represent the log-transformed series, we would denote the series
pictured in Figure 2–19 as (1 B)(1 B12) zt or as §§12zt to indicate that both first-order
and 12th-order differencing were applied. Since we are assuming an MA(1) model for
both the first difference process and the 12th-order difference process, the model we wish
to fit can be denoted §§12zt c (1 b1B)(1 b2B12)t, where the parameters b1 and
b2 are to be determined based on one of several fitting criteria. The exact values we obtain
for the parameters will depend on the optimization method we use, but generally all methods will yield similar values. Least squares is probably the most common method used,
but maximum likelihood and Bayesian methods have also been suggested. Using
XLSTAT, a program that contains an ARIMA forecasting module and is embedded in
Excel, we obtain the parameter values b1 0.333 and b2 0.544. The value of the constant c is small enough to be ignored. [These values differ slightly from those reported in
Box and Jenkins (1970), since the search algorithm used by XLSTAT differs from the one
used by Box and Jenkins.]
When forecasting using this model, it is convenient to write it out explicitly in the form
zt zt1 (zt12 zt13) t b1t1 b2(t12 b1t13).
Substituting for the parameter values and rearranging terms, the forecasting equation we
obtain for the log series is
zt zt1 zt12 zt13 t 0.333t1 0.544t12 0.181t13.
To forecast the original series, Dt, we apply the antilog to zt, namely, Dt exp(zt). Because of
the two levels of differencing, period 14 is the first period for which we can determine a forecast. In Figure 2–20, we show the original series starting at period 14 and the one-step-ahead
forecast using the preceding ARIMA model. Note how closely the ARIMA forecast tracks the
original series.
FIGURE 2–20
Observed versus predicted number of airline sales
700
Number of passengers
600
500
400
300
200
100
0
14 20 26 32 38 44 50 56 62 68 74 80 86 92 98 104 110 116 122 128 134 140
Period
Observed
Predicted
nah77856_ch02.indd 104
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 104
Snapshot Application
A SIMPLE ARIMA MODEL PREDICTS THE
PERFORMANCE OF THE U.S. ECONOMY
In years past, a very complex and very large regression
model known as the FRB-MIT-PENN (FMP) model (for
Federal Reserve Bank—Massachusetts Institute of
Technology—University of Pennsylvania) was used to
predict several basic measures of the U.S. economy. This
model required massive amounts of data and past history. Nelson (1972) employed the ARIMA methodology
outlined in this section to obtain predictors for many of
the same fundamental measures of the U.S. economy
considered in the FMP model. These include gross
national product, consumer’s expenditures on durable
goods, nonfarm inventory investment, and several others.
2.10 Box-Jenkins Models 105
Perhaps the most interesting case is the prediction of
the gross national product. The ARIMA model he
obtained is surprisingly simple:
zt zt 1 0.615(zt 1 zt 2) 2.76 t,
which is easily seen to be an AR(1) model with one level
of differencing. What is most impressive and surprising
is that the forecast errors obtained from this and
Nelson’s other ARIMA models had lower forecast errors
in predicting future values of these measures than did
the complex FMP model. Again, this points to the
power of these methods in providing accurate forecasts
in a variety of scenarios.
Box-Jenkins Modeling—A Critique
The preceding case study highlights the power of ARIMA models. However, one must
be aware that there are only a limited number of situations in which one would or could
use them. An important requirement is that one must have a substantial history of time
series in question. Typical recommendations vary between 50 and 100 observations of
past history, and that is for a nonseasonal series. When seasonality is present, the requirement is more severe. In some sense, each season of data is comparable to a single
observation.
In the operations context, one would be hard-pressed to find many applications where
that much history is available. Style goods, for example, are typically managed with a
very small amount of data, if any; for example, recall the Sport Obermeyer Snapshot
Application in this chapter. Even when enough data are available, it may not make sense
to invest the amount of time and energy required to develop an ARIMA model. Consider forecasting usage rates in a Wal-Mart store, for example, where one must manage
tens of thousands of SKUs. ARIMA models are most useful for forecasting economic
series with substantial history (such as GNP) or continuous processes (such as chemical
processes or stock market prices). In the latter case, one can conceivably choose a small
enough time bucket to generate any number of observations.
Another shortcoming of ARIMA models is that they are not easily updated. If the
problem parameters change, one must redo the analysis to find a new model, or at least
rerun the software to find new optimal values of the parameters.
Even with all of these shortcomings, Box-Jenkins methods are very powerful. In the
right situation, they can result in much more accurate short-term forecasts than any of
the other methods discussed in this chapter.
Problems for Section 2.10
38. Consider the white noise process data in Table 2–1. Enter this data into Excel and
generate the sample autocorrelations for lags of 1 to 10 periods, using the formula
for the sample autocorrelations (rk). Alternatively, if you have access to an ARIMA
computing module, generate these autocorrelations using your software.
nah77856_ch02.indd 105
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 105
106 Chapter Two Forecasting
2.11
Practical Considerations 105
39. Use the random number generator in Excel and the normal variate generator given
in Problem 5.34 to generate a sample path for an AR(1) process such as the ones
shown in Figure 2–14 for a1 0.7. Compute the sample autocorrelation function
for the process you generated, and check that the correlogram has the same structure as shown in Figure 2–15a.
40. Use the random number generator in Excel and the normal variate generator given
in Problem 5.34 to generate a sample path for an MA(1) process such as the ones
shown in Figure 2–16 for b1 0.7. Compute the sample autocorrelation function
for the process you generated, and check that the correlogram shows only one significant autocorrelation at a lag of one period. What does the value of this autocorrelation appear to be?
41. Use the random number generator in Excel and the normal variate generator given
in Problem 5.34 to generate a sample path for an ARMA(1,1) process. Use a1 0.8
and b1 0.6. Compute the sample autocorrelation function for this process.
42. Consider an ARIMA(2,1,1) process. Write out the resulting model using
a. Backshift notation with the backshift operator B.
b. Backshift notation with the operator §.
c. No backshift notation.
43. Consider an ARIMA(0,2,2) process. Write out the resulting model using
a. Backshift notation with the backshift operator B.
b. Backshift notation with the operator §.
c. No backshift notation.
44. Using back-substitution, show that an ARMA(1, 1) model may be written as either
an AR( q ) or an MA( q ) model.
45. Consider the seasonal time series pictured in Figure 2–8. What level of differencing would be required to make this series stationary?
46. The U.S. Federal Reserve in St. Louis stores a host of economic data on its Web
site at http://research.stlouisfed.org/fred2/categories/106. Download the following
time series from this Web site:
a. U.S. GNP.
b. Annual Federal Funds rate.
c. Consumer price index.
In each case, use a minimum of 50 data points over a period not including the
most recent 25 years. Graph the data points, and determine first the level of differencing that appears to be required. Once you have obtained a stationary
process, use the methods outlined in this section to arrive at an appropriate
ARIMA model for each case. Compare the most recent 25 years of observations
with the predictions obtained from your model using both one-step-ahead and
two-step-ahead forecasts.
2.11
PRACTICAL CONSIDERATIONS
Model Identification and Monitoring
Determining the proper model depends both on the characteristics of the history of observations and on the context in which the forecasts are required. When historical data
are available, they should be examined carefully in order to determine if obvious
nah77856_ch02.indd 106
17/12/14 6:47 PM
nah77856_ch02.qxd
106
12/31/07
4:21 PM
Page 106
Chapter Two Forecasting
2.11 Practical Considerations 107
patterns exist, such as trend or seasonal fluctuations. Usually, these patterns can be spotted by graphing the data. Statistical tests, such as significance of regression, can be used
to verify the existence of a trend, for example. Identifying complex relationships
requires more sophisticated methods. The sample autocorrelation function can reveal
intricate relationships that simple graphical methods cannot as we saw in Section 2.10.
Once a model has been chosen, forecasts should be monitored regularly to see if
the model is appropriate or if some unforeseen change has occurred in the series. As we
indicated, a forecasting method should not be biased. That is, the expected value of the
forecast error should be zero. In addition to the methods mentioned in Section 2.6, one
means of monitoring the bias is the tracking signal developed by Trigg (1964). Following
earlier notation, let et be the observed error in period t and et the absolute value of the
observed error. The smoothed values of the error and the absolute error are given by
Et � et � (1 � )Et�1,
Mt � et � (1 � )Mt�1.
The tracking signal is the ratio
E
Tt � ��t .
Mt
If forecasts are unbiased, the smoothed error Et should be small compared to the
smoothed absolute error Mt. Hence, a large value of the tracking signal indicates biased
forecasts, which suggest that the forecasting model is inappropriate. The value of Tt
that signals a significant bias depends on the smoothing constant . For example, Trigg
(1964) claims that a value of Tt exceeding 0.51 indicates nonrandom errors for a of .1. The tracking signal also can be used directly as a variable smoothing constant.
This is considered in Problem 55.
Simple versus Complex Time Series Methods
The literature on forecasting is voluminous. In this chapter we have touched only on
a number of fairly simple techniques. The reader is undoubtedly asking himself or
herself, do these methods actually work? The results from the literature suggest that the
simplest methods are often as accurate as sophisticated ones. Armstrong (1984)
reviews 25 years of forecasting case studies with the goal of ascertaining whether or
not sophisticated methods work better. In comparing the results of 39 case studies, he
found that in 20 cases sophisticated methods performed about as well as simple ones,
in 11 cases they outperformed simple methods, and in 7 cases they performed significantly worse.
A more sophisticated forecasting method is one that requires the estimation of a
larger number of parameters from the data. Trouble can arise when these parameters
are estimated incorrectly. To give some idea of the nature of this problem, consider a
comparison of simple moving averages and regression analysis for the following
series: 7, 12, 9, 23, 27. Suppose that we are interested in forecasting at the end of
period 5 for the demand in period 15 (that is, we require F5,15). The five-period
moving-average forecast made at the end of period 5 is 15.6, and this would be the
forecast for period 15. The least squares fit of the data is D̂t � 0.3 � 5.1t. Substituting t � 15, we obtain the regression forecast of 76.8. In Figure 2–21 we picture the
realization of the demand through period 15. Notice what has happened. The apparent
trend that existed in the first five periods was extrapolated to period 15 by the regression equation. However, there really was no significant trend in this particular
nah77856_ch02.indd 107
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 107
2.12
108 Chapter Two Forecasting
Overview of Advanced Topics in Forecasting 107
FIGURE 2–21
The difficulty with
long-term forecasts
100
90
80
Regression
forecast
Demand
70
60
50
40
30
Movingaverage
forecast
20
10
0
1
2
3 4
Baseline
period
5
6
7
8
9
10 11 12 13 14 15
Time
case. The more complex model gave significantly poorer results for the long-term
forecast.
There is some evidence that the arithmetic average of forecasts obtained from different methods is more accurate than a single method (see Makridakis and Winkler, 1983).
This is perhaps because often a single method is unable to capture the underlying signal
in the data and different models capture different aspects of the signal. (See a discussion
of this phenomenon following Armstrong, 1984.)
What do these observations tell us about the application of forecasting techniques
to production planning? At the aggregate level of planning, forecast accuracy is
extremely important and multiple-step-ahead forecasts play an integral role in the
planning of workforce and production levels. For that reason, blind reliance on time series methods is not advised at this level. At a lower level in the system, such as routine
inventory management for spare parts, the use of simple time series methods such as
moving averages or exponential smoothing makes a great deal of sense. At the individual item level, short-term forecasts for a large number of items are required, and
monitoring the forecast for each item is impractical at best. The risk of severe errors is
minimized if simple methods are used.
2.12
OVERVIEW OF ADVANCED TOPICS IN FORECASTING
Simulation as a Forecasting Tool
Computer simulation is a powerful technique for tackling complex problems. A computer simulation is a description of a problem reduced to a computer program. The program is designed to re-create the key aspects of the dynamics of a real situation. When
a problem is too complex to model mathematically, simulation is a popular alternative. By rerunning the program under different starting conditions and/or different
scenarios, one can, by a kind of trial-and-error process, discover the best strategy for
managing a system.
nah77856_ch02.indd 108
17/12/14 6:47 PM
nah77856_ch02.qxd
108
12/31/07
4:21 PM
Chapter Two Forecasting
Page 108
2.12 Overview of Advanced Topics in Forecasting 109
Simulation is a common tool for modeling manufacturing planning problems such
as complex material flow problems in the plant. It is less commonly used as a forecasting tool. Compaq Computer, a successful producer of personal computers based
in Houston, Texas, has experimented with a powerful simulation-based forecasting
tool for assisting with the process of new product introductions (McWilliams, 1995).
The program recommends the optimal timing and pricing of new product introductions by incorporating forecasts of component availability and price changes, fluctuating demand for a given feature or price, and the impact of rival models.
Using this tool, Compaq decided to delay announcement of several new Pentiumbased models in late 1994. This strategy “went against everything the company believed.” Compaq’s basic strategy had always been to be a technology leader, but its
forecasting tool suggested that corporate customers were not quite ready to switch to
Pentium-based machines at the end of 1994. The strategy proved to be very profitable
for Compaq, which subsequently posted record earnings.
Forecasting Demand in the Presence of Lost Sales
Retailers rely heavily on forecasting. Basic items (items that don’t change appreciably
from season to season, such as men’s dress shirts) generally have substantial sales
history, arguing for the use of time series methods to forecast demand. However, there
is an important difference between what is observed and what one wants to forecast.
The goal is to forecast demand, but one only observes sales. What’s the difference?
Suppose a customer wants to buy a blouse in a certain size and color and finds it’s not
available on the shelf? What will she do? Perhaps she will place a special order with a
salesperson, but, more likely, she will just leave the store and try to find the product
somewhere else. This is known as a lost sale. The difficulty is that most retailers have
no way to track lost sales. Thus, they observe sales but need to estimate demand.
As an example, consider an item that is restocked to 10 units at the beginning of each
week. Suppose that over the past 15 weeks the sales history for the item was 7, 5, 10, 10,
8, 3, 6, 10, 10, 9, 5, 0, 10, 10, 4. Consider those weeks in which sales were 10 units. What
were the demands in those weeks? The answer is that we don’t know. We only know that
it was at least 10. If you computed the sample mean and sample variance of these numbers, they would underestimate the true mean and variance of demand.
How does one go about forecasting demand in this situation? In the parlance of
classical statistics, this is known as a censored sample. That means that we know the
values of demand for only a portion of the sample. For the other portion of the sample,
we know only a lower bound on the demand. Special statistical methods that incorporate
censoring give significantly improved estimates of the population mean and variance in
this case. These methods can be embedded into sequential forecasting schemes, such as
exponential smoothing, to provide significantly improved forecasts.
Nahmias (1994) considered the problem of forecasting in the presence of lost sales
when the true demand distribution was normal. He compared the method of maximum
likelihood for censored samples and a new method, either of which could be incorporated into exponential smoothing routines. He also showed that both of these methods
would result in substantially improved forecasts of both the mean and the variation of
the demand.
To see how dramatic this difference can be, consider a situation in which the true
weekly demand for a product is a normal random variable with mean 100 and standard deviation 30. Suppose that items are stocked up to 110 units at the start of each
week. Exponential smoothing is used to obtain two sets of forecasts: the first accounts for lost sales (includes censoring) and the second does not (does not including
censoring). Figures 2–22 and 2–23 show the estimators for the mean and the standard
nah77856_ch02.indd 109
17/12/14 6:47 PM
nah77856_ch02.qxd
12/31/07
4:21 PM
Page 109
2.12
110 Chapter Two Forecasting
Overview of Advanced Topics in Forecasting 109
FIGURE 2–22
Tracking the mean when lost sales are present
115
110
105
Mean
100
95
90
85
80
75
11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40
Period
Mean value
Censored estimate
Uncensored estimate
FIGURE 2–23
Tracking the standard deviation when lost sales are present
40
Standard deviation
35
30
25
20
15
10
11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40
Period
Standard deviation
Censored estimate
Uncensored estimate
deviation, respectively, with and without censoring. Notice the severe low bias when lost
sales are ignored in both cases. That means that by not correctly accounting for the
difference between sales and demand, one underestimates both the mean and the variance of the demand. Since both the mean and the variance of demand are inputs for
determining optimal stocking levels, these levels could be severely underestimated.
nah77856_ch02.indd 110
17/12/14 6:47 PM
nah77856_ch02.qxd
110
12/31/07
4:21 PM
Chapter Two Forecasting
2.13
Page 110
2.13 Linking Forecasting and Inventory Management 111
LINKING FORECASTING AND INVENTORY MANAGEMENT
Inventory control under demand uncertainty will be treated in detail in Chapter 5. In
practice, inventory management and demand forecasting are closely linked. The forecasting method could be any one of the methods discussed in this chapter. One of the
inputs required for inventory control models is the distribution of the demand over a
period or over an order replenishment lead time. Is there a link between the distribution
of forecast errors and the distribution of demand? The answer is that there is such
a link. The forecast error distribution plays a key role in the correct application of
inventory models in practice.
The first issue is to decide on the appropriate form of the distribution of demand.
Most commercial systems assume that the demand distribution is normal. That means
we need only estimate the mean and the standard deviation to specify the entire
distribution. (Statistical methods, known as goodness-of-fit techniques, can be applied
to test the accuracy of the normality assumption.) Whether or not the inventory system
is linked to a forecasting system, we must have a history of observations of demand to
obtain statistical estimates of the mean and variance. (When no history of demand
exists, personal judgment must be substituted for statistical estimation. Methods for
aggregating subjective judgment are discussed in Section 2.3.)
In the beginning of Chapter 5, we discuss how one would estimate the mean and
the variance of demand directly from a history of demand observations. In practice,
however, we don’t generally use a long history of observations, because we believe
that the underlying demand distribution does not remain constant indefinitely. That
is why we adjust the N for moving averages or the for exponential smoothing to
balance stability and responsiveness. If that is the case, what is the appropriate variance estimator one should use?
In Appendix 2–A, we show that for simple moving averages of order N, the variance
of forecast error, 2e, is given by
N�1
2e � 2 ��
N
and for exponential smoothing the variance of forecast error is
2
2e � 2 �� .
2�
Notice that in both cases, the value of 2e exceeds the value of 2. Also notice that
as N gets large and as gets small, the values of e and grow close. These cases
occur when one uses the entire history of demand observations to make a forecast.
In Chapter 5 one of the inputs needed to determine safety stocks for inventory is the
distribution of demand. The problem is, if we have estimators for both e and ,
which should be used as the standard deviation estimator for setting safety stocks?
The obvious answer is that we should use the estimator for , since this represents
the standard deviation of demand. The correct answer, however, is that we should
use e. The reason is that the process of forecasting introduces sampling error into the
estimation process, and this sampling error is accounted for in the value of e. The
forecasting error variance is higher than the demand variance because the forecast
is based on only a limited portion of the demand history.
We also can provide an intuitive explanation. If a forecast is used to estimate the
mean demand, we keep safety stocks in order to protect against the error in this forecast.
nah77856_ch02.indd 111
17/12/14 6:47 PM
112 Chapter Two Forecasting
Hence, the distribution of forecast errors is more relevant than the distribution of
­demands. This is an important practical point that has been the source of much confusion in the literature. We will return to this issue in Chapter 5 and discuss its relevance
in the context of safety stock calculations.
Most inventory control systems use the method suggested by R. G. Brown (1959
and 1962) to estimate the value of se. (In fact, it appears that Brown was the first
to recognize the importance of the distribution of forecast errors in inventory management applications.) The method requires estimating the MAD of forecast errors
using exponential smoothing. This is accomplished using the smoothing equation
MADt 5 a|Ft 2 Dt| 1 (1 2 a)MADt21
The MAD is converted to an estimate of the standard deviation of forecast error by
multiplying by 1.25. That is, the estimator for se obtained at time t is
ŝe 5 1.25 MADt.
A small value of a, generally between 0.1 and 0.2, is used to ensure stability in the MAD
estimator. This approach to estimating the MAD works for any of the forecasting methods discussed in this chapter. Safety stocks are then computed using this estimator for se.
Case Study. Predicting the Growth of Facebook
Facebook, founded by Mark Zuckerberg in 2004, is the quintessential social network and
one of the true phenomena of the early 21st century. While Myspace, another social n­ etwork,
was founded a year earlier than Facebook, it never enjoyed the kind of success experienced
by Facebook. Consider the problem of forecasting the number of Facebook users in 2013
based on data up until 2011. The following chart tracks the numbers of Facebook users from
the year of its founding in 2004 to early 2011:
DateUsers (in millions)Month Number
12/04
12/05
12/06
4/07
10/07
08/08
01/09
02/09
04/09
07/09
09/09
12/09
02/10
07/10
09/10
01/11
02/11
1
5.5
12
20
50
100
150
175
200
250
300
350
450
500
550
600
650
12
24
36
40
46
56
61
62
64
67
69
72
74
79
81
85
86
Note that the dates are not evenly spaced. One way to graph this data, which will also
be useful in analyzing it, is to convert the dates to numbers of months elapsed from an
arbitrary starting point. If we define month 1 as January 2004, then the dates in column
1 translate to the number of months elapsed in column 3.
nah77856_ch02.indd 112
17/12/14 6:47 PM
2.13 Linking Forecasting and Inventory Management 113
Treating numbers of months as the independent variable, the graph of numbers of
users versus months elapsed appears in the graph below:
700
600
500
400
300
200
100
0
0
20
40
60
80
100
The graph seems to show exponential growth. To test this hypothesis, we consider
graphing the natural logarithm of the numbers of users versus elapsed numbers of
months. Doing so results in the following graph:
7
6
5
4
3
2
1
0
nah77856_ch02.indd 113
0
20
40
60
80
100
17/12/14 6:47 PM
114 Chapter Two Forecasting
The fact that the graph of the natural logs is almost linear means that the assumption of
exponential growth is an accurate one. Fitting a simple linear regression to this curve
results in the following:
SUMMARY OUTPUT
Regression Statistics
Multiple R
R Square
Adjusted R Square
Standard Error
Observations
0.987306
0.974772
0.973091
0.307178
17
ANOVA
df
Regression
Residual
Total
Coefficients
Intercept
1
15
16
Standard
Error
t Stat
SS
MS
54.68893
1.415375
56.10431
P-value
54.68893
0.094358
Lower 95%
F
Significance F
579.5877578
2.11649E-13
Upper 95% Lower 95.0%
Upper
95.0%
0.4425 0.22592 1.95867 0.069013361 0.924037859 0.039035397 0.924037859 0.039035397
X Variable 1 0.086085 0.00357 24.07463 2.11649E-13 0.078463009 0.093706009 0.078463009 0.093706009
This Excel output shows that the logs follow a straight line relationship very closely. In
fact, it’s rare to see an R square value of more than .97, so this fit is extremely strong.
Let’s see what numbers of users would be projected by this model.
The regression model is y = a + bx where y represents the logarithm of the number of users, a is the intercept and b is the slope. The independent variable, x, is the
number of months elapsed since January 2004. The regression output indicates that the
least squares estimators are a = -.4425 and b = .086085. Let’s consider how this model
would be used for forecasting numbers of Facebook users. Consider the forecast for
January 2012. This corresponds to month 97. Setting x = 97 gives y = 7.90785. Since
y = ln (y), it follows that y = exp(y ) = exp (7.90785) = 2718.53 millions of users. That
is, if Facebook’s growth continues at its current rate, there should be approximately
2.7 billion users as soon as January 2012. This seems unlikely, but possible. Let’s
consider the prediction for January 2014. This month corresponds to x = 121, which
results in y = 9.34 and a predicted number of users of more than 11 billion! Obviously,
this prediction is absurd, as it is almost twice as many people as there are on earth.
The conclusion is that exponential growth cannot continue indefinitely if there are
finite resources involved (in this case, the total number of computer users on the Earth).
So the question remains, how should one forecast the growth of Facebook users? Clearly
the model of continued exponential growth is unreasonable. What is likely is that the
observed exponential growth is only the first phase of a more complex model. But
what would be an appropriate model in this case? The answer is that at this point we’re
not sure, but based on past experience with company growth curves, it is reasonable
to postulate a logistic curve. The mathematical form of the standard logistic curve is
1
. This results in the curve pictured below.
P(t) =
1 + e-t
¬
¬
¬
¬
¬
¬
nah77856_ch02.indd 114
17/12/14 6:47 PM
2.13 Linking Forecasting and Inventory Management 115
FPO
The goal is to try to provide the best fit of this curve to the original Facebook data.
Since we’re not sure where the point of inflection occurs, our estimate at this stage is
probably not going to be very accurate. However, we can do a lot better than simply
extrapolating from an exponential growth curve. In order to fit a logistic curve to a
set of data, we need to know a few things. First, where in the data set is the inflection
point? Second, what is likely to be the value of the asymptote (which is one for the
standard curve)? And finally, what is the appropriate time scale?
Consider estimating the asymptote. This is probably not going to be very accurate
since we don’t know where the inflection point is. From the original graph, it does appear that the inflection point occurs somewhere near month 80. The number of users in
month 80 is approximately 525 million. This would imply a value of the asymptote of
approximately (2)(525 million) = 1.05 billion. This would be our best guess of where
the total numbers of Facebook users would top out in the coming years.
We also need to transform the time scale. If we assume in our original curve that the first
half of the logistic curve corresponds to months 20 to 80, this would be the same as 5 to 0
in the standard logistic curve above. The reader should satisfy him or herself that this would
correspond to the transformation of the time scale following the equation u = 12*t + 80.
(Check the values of u and t at 5, 0, and 5 to verify this transformation.) Furthermore, if
the asymptote is K rather than 1, we scale the function by multiplying by K. Applying these
two transformations, the appropriate logistic curve to fit our original data would be
P(u) =
K
1 + exp( -(u - 80) / 12))
where u corresponds to number of elapsed months, and K = 1.05 × 109. Consider
the forecast for January 2012 based on this model. Substituting u = 97 gives a forecast of P(u)= 0.845 billion users. This certainly seems reasonable. According to the
­logistic model, we would expect that the number of users will be close to the postulated
­maximum of 1.05 billion by month 140, which corresponds to August of 2015.
nah77856_ch02.indd 115
17/12/14 6:48 PM
116 Chapter Two Forecasting
In late 2013 Facebook reported that there were 1.158 billion users. If true, this implies
that our model predicted low. However, sources claimed that approximately 5 percent
of those may have been duplicate accounts, while 1.3 percent may have been accounts
that were improperly classified by users, and 0.9 percent may have been fake accounts
(Cohen, 2013). That means the true value is likely to be closer to 1.07 billion, not far
from the model’s asymptote.
2.14 Historical Notes and Additional Topics
Forecasting is a rich area of research. Its importance in business applications cannot be
overstated. The simple time series methods discussed in this chapter have their roots in
basic statistics and probability theory. The method of exponential smoothing is g­ enerally
credited to R. G. Brown (1959 and 1962), who worked as a consultant to A. D. Little at
the time. Although not grounded in basic theory, exponential smoothing is probably one
of the most popular forecasting methods used today. Brown was also the first to recognize the importance of the forecast error distribution and its i­ mplications for inventory
management. However, interest in using statistical methods for forecasting goes back to
the turn of the century or before. (See, for example, Yule, 1926.)
Not discussed in this chapter is forecasting of time series using spectral analysis and
state space methods. These methods are highly sophisticated and require substantial
structure to exist in the data. They often rely on the use of the autocorrelation function,
and in that sense are similar conceptually to Box-Jenkins methods discussed briefly in
Section 2.10. The groundbreaking work in this area is due to Norbert Wiener (1949)
and Rudolph Kalman (1960). However, there have been few applications of these
methods to forecasting economic time series. Most applications have been in the area
of signal processing in electrical engineering. [Davenport and Root (1958) provide a
good summary of the fundamental concepts in this area.]
The Kalman filter is a type of exponential smoothing technique in which the value of
the smoothing constant changes with time and is chosen in some sort of optimal fashion.
The idea of adjusting the value of the smoothing constant based on some m
­ easure of
prior performance of the model has been used in a number of ad hoc ways. A typical
­approach is the one suggested by Trigg and Leach (1967), which requires the calculation of the tracking signal. The tracking signal is then used as the value of the smoothing
constant for the next forecast. The idea is that when the tracking signal is large, it suggests that the time series has undergone a shift; a larger value of the smoothing constant
should be more responsive to a sudden shift in the underlying signal. Other methods
have also been suggested. We have intentionally not included an explicit d­ iscussion of
adaptive response rate methods for one reason: there is little e­ vidence that they work in
the context of predicting economic time series. Most studies that compare the effectiveness of different forecasting methods for many different series show no ­advantage for
adaptive response rate models. (See, for example, ­Armstrong, 1984.) The method of
Trigg and Leach is discussed in more detail in P
­ roblem 55.
With the sixth edition, we have included a reasonably self-contained discussion of
Box-Jenkins ARIMA models. The basic ideas behind these methods, such as the autocorrelation structure of a process, go back many years. However, Box and Jenkins
(1970) were the first to put these ideas together into a comprehensive step-by-step
­approach for building ARIMA models for short-term forecasting. Readers may find
nah77856_ch02.indd 116
17/12/14 6:48 PM
nah77856_ch02.qxd
nah77856_ch02.qxd
1/17/08 1/17/08
3:14 PM 3:14
PagePM113Page 113
Additional Problems
Additional
onProblems
Problems
Forecasting
on Forecasting 117
Forecasting
113
113
Additional
on
their book
their
daunting,
book daunting,
as they focus
as they
onfocus
issuesonofissues
mathematical
of mathematical
concern.concern.
For the reader
For the reader
seeking aseeking
more comprehensive
a more comprehensive
coveragecoverage
of ARIMA
of ARIMA
models atmodels
a levelatconsistent
a level consistent
with
with
ours, a good
ours,starting
a goodpoint
starting
is the
point
textisby
theMakridakis,
text by Makridakis,
Wheelright,
Wheelright,
and Hyndman
and Hyndman
(1998). (1998).
2.15 Summary
2.15 Summary
This chapter
This provided
chapter provided
an introduction
an introduction
to a number
to a
of
number
the more
of the
popular
more time
popular
series
time series
forecasting
forecasting
techniques
techniques
as well as
as awell
brief
as discussion
a brief discussion
of other of
methods,
other methods,
includingincluding
the
the
Delphi method
Delphi and
method
causal
andmodels.
causal A
models.
moving-average
A moving-average
forecast is
forecast
obtained
is obtained
by computby computing the arithmetic
ing the arithmetic
average of
average
the N of
most
therecent
N most
observations
recent observations
of demand.
of demand.
An exponential
An exponential
smoothing
smoothing
forecast forecast
is obtained
is obtained
by computing
by computing
the weighted
the weighted
average average
of the current
of the current
observation
observation
of demand
of and
demand
the most
and the
recent
most
forecast
recent of
forecast
that demand.
of that demand.
The weight
Theapplied
weight applied
to the current
to the observation
current observation
is and isthe weight
and theapplied
weight to
applied
the last
to forecast
the last (that
forecast
is, the
(that is, the
past observations)
past observations)
is 1 .
is Small
1 .N Small
and large
N and
result
large in result
responsive
in responsive
forecasts,forecasts,
and
and
large N and
largesmall
N and
result
small in result
stable inforecasts.
stable forecasts.
AlthoughAlthough
the two methods
the two methods
have similar
have similar
properties,
properties,
exponential
exponential
smoothing
smoothing
is generally
is generally
preferredpreferred
because itbecause
requires
it storing
requiresonly
storing only
the previous
the previous
forecast, forecast,
whereas whereas
moving averages
moving averages
requires requires
storing the
storing
last the
N demand
last N demand
observations.
observations.
When there
When
is a there
trendisinathe
trend
series,
in the
both
series,
moving
bothaverages
moving averages
and exponential
and exponential
smoothing
smoothing
lag
lag
behind the
behind
trend.the
Wetrend.
discussed
We discussed
two timetwo
series
time
techniques
series techniques
that are designed
that are designed
to track the
to track the
trend. One
trend.
is regression
One is regression
analysis, analysis,
which uses
which
leastuses
squares
leasttosquares
fit a straight
to fit aline
straight
to theline
data,
to the data,
and the other
and the
is Holt’s
other is
method,
Holt’s which
method,is which
a type isofadouble
type ofexponential
double exponential
smoothing.
smoothing.
Holt’s Holt’s
method has
method
the advantage
has the advantage
that forecasts
that forecasts
are easierare
to easier
updatetoasupdate
new demand
as new observations
demand observations
become available.
become available.
We also We
discussed
also discussed
techniques
techniques
for seasonal
for seasonal
series. We
series.
employed
We employed
classical classical
decomposidecomposition to show
tion how
to show
simple
howmoving
simpleaverages
moving averages
could be could
used to
beestimate
used to estimate
seasonal seasonal
factors and
factors and
obtain theobtain
deseasonalized
the deseasonalized
series when
series
there
when
is athere
trendisand
a trend
showed
and how
showed
seasonal
how seasonal
factors factors
could be could
estimated
be estimated
quickly when
quickly
there
when
is no
there
trend.
is no
Thetrend.
extension
The extension
of Holt’s of
method
Holt’stomethod
deal to deal
with seasonal
with seasonal
problems,problems,
called Winters’s
called Winters’s
method, is
method,
a type of
is atriple
type exponential
of triple exponential
smoothing
smoothing
technique.
technique.
Section 2.12
Section
provided
2.12 provided
a brief overview
a brief overview
of advanced
of advanced
methods methods
that are that
beyond
are the
beyond the
scope of scope
this text.
of this
Thetext.
finalThe
section
finaldiscussed
section discussed
the relationship
the relationship
between between
forecasting
forecasting
and
and
inventoryinventory
control. The
control.
key The
point,
key
which
point,will
which
be elaborated
will be elaborated
on in Chapter
on in 5,
Chapter
is that5,the
is that the
standard standard
deviationdeviation
of forecast
of error
forecast
is the
error
appropriate
is the appropriate
measure measure
of variation
of variation
for computfor computing safetying
stocks.
safety stocks.
Additional
Additional 47.
Problems
Problems
on
on
Forecasting
Forecasting
John47.
Kittle,
JohnanKittle,
independent
an independent
insuranceinsurance
agent, uses
agent,
a five-year
uses a five-year
moving average
moving to
average to
forecast the
forecast
number
the of
number
claimsofmade
claims
in made
a single
in year
a single
for one
year of
forthe
onelarge
of the
insurance
large insurance
companies
companies
he sells for.
he sells
He has
for.just
He discovered
has just discovered
that a clerk
thatina his
clerk
employ
in hisincorrectly
employ incorrectly
entered the
entered
number
the of
number
claimsofmade
claims
four
made
years
four
agoyears
as 1,400
ago as
when
1,400
it should
when ithave
should have
been 1,200.
been 1,200.
a. What a.adjustment
What adjustment
should Mr.
should
Kittle
Mr.make
Kittlein make
next year’s
in nextforecast
year’s forecast
to take into
to take into
account the
account
corrected
the corrected
value of value
the number
of the of
number
claimsoffour
claims
yearsfour
ago?
years ago?
b. Suppose
b. Suppose
that Mr.that
Kittle
Mr. used
Kittlesimple
used exponential
simple exponential
smoothing
smoothing
with with
.2 .2
instead ofinstead
moving
ofaverages
moving averages
to determine
to determine
his forecast.
his forecast.
What adjustment
What adjustment
is now re-is now required inquired
next year’s
in next
forecast?
year’s forecast?
(Note that(Note
you do
thatnot
you
need
do not
to know
need the
to know
valuethe
of the
value of the
forecast for
forecast
next year
for next
in order
year to
in solve
order this
to solve
problem.)
this problem.)
nah77856_ch02.indd 117
17/12/14 6:48 PM
nah77856_ch02.qxd
1/17/08
3:14 PM
Page 114
nah77856
118 Chapter Two Forecasting
48. A method of estimating the MAD discussed in Section 2.13 recomputes it each time a
new demand is observed according to the following formula:
MADt � et � (1 � ) MADt�1.
Consider the one-step-ahead forecasts for aircraft engine failures for quarters 2
through 8 obtained in Example 2.3. Assume an initial value of the MAD � 50 in
period 1. Using the same value of , what values of the MAD does this method give
for periods 2 through 8? Discuss the advantages and disadvantages of this approach
vis-à-vis direct computation of the MAD.
49. Herman Hahn is attempting to set up an integrated forecasting and inventory control
system for his hardware store, Hahn’s Hardware. When Herman indicates that outdoor
lights are a seasonal item on the computer, he is prompted by the program to input the
seasonal factors by quarter.
Unfortunately, Herman has not kept any historical data, but he estimates that firstquarter demand for the lights is about 30 percent below average, the second-quarter
demand about 20 percent below average, third-quarter demand about average, and
fourth-quarter demand about 50 percent above average. What should he input for the
seasonal factors?
50. Irwin Richards, a publisher of business textbooks, publishes in the areas of management, marketing, accounting, production, finance, and economics. The president of the
firm is interested in getting a relative measure of the sizes of books in the various
fields. Over the past three years, the average numbers of pages of books published in
each area were
Management
Marketing
Accounting
Production
Finance
Economics
Year 1
Year 2
Year 3
835
620
440
695
380
1,220
956
540
490
680
425
1,040
774
575
525
624
410
1,312
Using the quick and dirty methods discussed in Section 2.9, find multiplicative factors
for each area giving the percentage above or below the mean number of pages.
51. Over a two-year period, the Topper Company sold the following numbers of lawn
mowers:
Month:
Sales:
Month:
Sales:
1
2
3
4
5
6
7
8
9
10
11
12
238
220
195
245
345
380
270
220
280
120
110
85
13
14
15
16
17
18
19
20
21
22
23
24
135
145
185
219
240
420
520
410
380
320
290
240
a. In column A input the numbers 1 to 24 representing the months and in
column B the observed monthly sales. Compute the three-month moving-average
forecast and place this in the third column. Be sure to align your forecast with the
period for which you are forecasting (the average of sales for months 1, 2, and 3
should be placed in row 4; the average of sales for months 2, 3, and 4 in row 5; and
so on.) In the fourth column, compute the forecast error for each month in which
you have obtained a forecast.
nah77856_ch02.indd 118
17/12/14 6:48 PM
nah77856_ch02.qxd
1/17/08
3:14 PM
Page 115
Additional Problems on Forecasting 119
b. In columns 5, 6, and 7 compute the absolute error, the squared error, and the
absolute percentage error. Using these results, find the MAD, MSE, and MAPE for
the MA(3) forecasts for months 4 through 24.
c. Repeat parts (a) and (b) for six-month moving averages. (These calculations
should appear in columns 8 through 12.) Which method, MA(3) or MA(6), was
more accurate for these data?
52. Repeat the calculations in Problem 51 using simple exponential smoothing, and allow
the smoothing constant to be a variable. That is, the smoothing constant should be a
cell location. By experimenting with different values of , determine the value that
appears to minimize the
a. MAD
b. MSE
c. MAPE
Assume that the forecast for month 1 is 225.
53. Baby It’s You, a maker of baby foods, has found a high correlation between the
aggregate company sales (in $100,000) and the number of births nationally the
preceding year. Suppose that the sales and the birth figures during the past eight
years are
Year
Sales (in $100,000)
U.S. births (in millions)
1
2
3
4
5
6
7
8
6.1
2.9
6.4
3.4
8.3
3.5
8.8
3.1
5.1
3.8
9.2
2.8
7.3
4.2
12.5
3.7
a. Assuming that U.S. births represent the independent variable and sales the
dependent variable, determine a regression equation for predicting sales
based on births. Use years 2 through 8 as your baseline. (Hint: You will require
the general regression formulas appearing in Appendix 2–B to solve this
problem.)
b. Suppose that births are forecasted to be 3.3 million in year 9. What forecast for
sales revenue in year 10 do you obtain using the results of part (a)?
c. Suppose that simple exponential smoothing with .15 is used to predict the
number of births. Use the average of years 1 to 4 as your initial forecast for period 5,
and determine an exponentially smoothed forecast for U.S. births in year 9.
d. Combine the results in parts (a), (b), and (c) to obtain a forecast for the sum of total
aggregate sales in years 9 and 10.
54. Hy and Murray are planning to set up an ice cream stand in Shoreline Park, described
in Problem 28. After six months of operation, the observed sales of ice cream (in
dollars) and the number of park attendees are
Month
Ice cream sales
Park attendees
nah77856_ch02.indd 119
1
2
3
4
5
325
880
335
976
172
440
645
1,823
770
1,885
6
950
2,436
17/12/14 6:48 PM
nah77856_ch02.qxd
1/17/08
3:15 PM
nah77856
Page 116
120 Chapter Two Forecasting
a. Determine a regression equation treating ice cream sales as the dependent variable
and time as the independent variable. Based on this regression equation, what
should the dollar sales of ice cream be in two years (month 30)? How confident are
you about this forecast? Explain your answer.
b. Determine a regression equation treating ice cream sales as the dependent variable
and park attendees as the independent variable. (Hint: You will require the general
regression equations in Appendix 2–B in order to solve this part.)
c. Suppose that park attendance is expected to follow a logistic curve (see Figure 2–24).
The park department expects the attendance to peak out at about 6,000
attendees per month. Plot the data of park attendees by month and “eyeball”
a logistics curve fit of the data using 6,000 as your maximum value. Based on your
curve and the regression equation determined in part (b), predict ice cream sales
for months 12 through 18.
55. A suggested method for determining the “right” value of the smoothing constant in
exponential smoothing is to retrospectively determine the value that results in the
minimum forecast error for some set of historical data. Comment on the appropriateness of this method and some of the potential problems that could result.
56. Lakeroad, a manufacturer of hard disks for personal computers, was founded in 1981
and has sold the following numbers of disks:
Year
Number Sold
(in 000s)
Year
Number Sold
(in 000s)
1981
1982
1983
1984
0.2
4.3
8.8
18.6
1985
1986
1987
1988
34.5
68.2
85.0
58.0
FIGURE 2–24
Number of
attendees
Logistic curve (for
Problem 52)
6,000
Time
nah77856_ch02.indd 120
17/12/14 6:48 PM
nah77856_ch02.qxd
1/17/08
3:15 PM
Page 117
Additional Problems on Forecasting 121
a. Suppose the firm uses Holt’s method for forecasting sales. Assume S0 � 0 and
G0 � 8. Using � .2 and � .2, find one-step-ahead forecasts for 1982 through
1989 and compute the MAD and MSE for the forecasts during this period. What
is the sales forecast for the year 2000 made at the end of 1988? Based on the
results of 1988, why might this forecast be very inaccurate?
b. By experimenting with various values of and , determine the values of the
smoothing constants that appear to give the most accurate forecasts.
57. Trigg and Leach (1967) suggest the following adaptive response-rate exponential
smoothing method. Along with smoothing the original series, also smooth the error et
and the absolute error et according to the equations
Et � et � (1 � )Et�1,
Mt � et � (1 � )Mt�1
and define the smoothing constant to be used in forecasting the series in period
t as
t �
2
Et
2.
Mt
The forecast made in period t for period t � 1 is obtained by the usual exponential smoothing equation, using t as the smoothing constant. That is,
Ft�1 � t Dt � (1 � t)Ft.
The idea behind the approach is that when E t is close in magnitude to Mt, it
suggests that the forecasts are biased. In that case, a larger value of the smoothing constant results that makes the exponential smoothing more responsive to sudden changes
in the series.
a. Apply the Trigg-Leach method to the data in Problem 22. Using the MAD, compare the accuracy of these forecasts with the simple exponential smoothing forecasts obtained in Problem 22. Assume that E1 � e1 (the observed error in January)
and M1 � e1. Use � .1.
b. For what types of time series will the Trigg-Leach adaptive response rate give more
accurate forecasts, and under what circumstances will it give less accurate forecasts? Comment on the advisability of using such a method for situations in which
forecasts are not closely monitored.
58. The owner of a small brewery in Milwaukee, Wisconsin, is using Winters’s method
to forecast his quarterly beer sales. He has been using smoothing constants of
� .2, � .2, and � .2. He has currently obtained the following values of the
various slope, intercept, and seasonal factors: S10 � 120, G10 � 14, c10 � 1.2,
c9 � 1.1, c8 � .8, and c7 � .9.
a. Determine the forecast for beer sales in quarter 11.
b. Suppose that the actual sales turn out to be 128 in quarter 11. Find S11 and G11, and
find the updated values of the seasonal factors. Also determine the forecast made at
the end of quarter 11 for quarter 13.
nah77856_ch02.indd 121
17/12/14 6:48 PM
nah77856_ch02.qxd
1/17/08
3:15 PM
nah77856
Page 118
122 Chapter Two Forecasting
59. The U.S. gross national product (GNP) in billions of dollars during the period 1964 to
1984 was as follows:
Year
GNP
Year
GNP
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
649.8
705.1
772.0
816.4
892.7
963.9
1,015.5
1,102.7
1,212.8
1,359.3
1,472.8
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1,598.4
1,782.8
1,990.5
2,249.7
2,508.2
2,732.0
3,052.6
3,166.0
3,401.6
3,774.7
Source: Economic Report of the President, February 1986.
a. Use Holt’s method to predict the GNP. Determine a regression fit of the data for
the period 1964 to 1974 to estimate the initial values of the slope and intercept.
(Hint: If you are doing the regression by hand, transform the years by subtracting
1963 from each value to make the calculations less cumbersome.) Using Holt’s
method, determine forecasts for 1975 to 1984. Assume that � .2 and � .1.
Compute the MAD and the MSE of the one-step-ahead forecasts for the period
1975 to 1984.
b. Determine the percentage increase in GNP from 1964 to 1984 and graph the
resulting series. Use a six-month moving average and simple exponential smoothing with � .2 to obtain one-step-ahead forecasts of this series for the period
1975 to 1984. (Use the arithmetic average of the observations from 1964 to
1974 to initialize the exponential smoothing.)
In both cases (i.e., MA and ES forecasts), convert your forecasts of the percentage increase for the following year to a forecast of the GNP itself and compute
the MAD and the MSE of the resulting forecasts. Compare the accuracy of these
methods with that of part (a).
c. Discuss the problem of predicting GNP. What methods other than the ones used in
parts (a) and (b) might give better predictions of this series?
Appendix 2–A
Forecast Errors for Moving Averages
and Exponential Smoothing
The forecast error et is the difference between the forecast for period t and the actual
demand for that period. In this appendix we will derive the distribution of the forecast error
for both moving averages and exponential smoothing.
nah77856_ch02.indd 122
17/12/14 6:48 PM
nah77856_ch02.qxd
1/17/08
3:15 PM
Page 119
Appendix 2–A 123
The demand is assumed to be generated by the process
Dt � � t,
where t is normal with mean zero and variance 2.
CASE 1. MOVING AVERAGES
Consider first the case in which forecasts are generated by moving averages. Then the forecast error is et � Ft � Dt, where Ft is given by
Ft �
1
N
a Di.
t�1
i�t�N
It follows that
a E(Di) � E(Dt) � (1N)(N) � � 0.
t�1
E(Ft � Dt) � (1N)
i�t�N
This proves that when demand is stationary, moving-average forecasts are unbiased.
Also,
Var(Ft � Dt) � Var(Ft) � Var(Dt)
� (1N2)
a Var(Di) � Var(Dt)
t�1
i�t�N
2
� (1N 2)(N ) � 2
� 2(1 � 1N) � 2[(N � 1)N].
It follows that the standard deviation of the forecast error, e, is
e � N�1
.
A N
This is the standard deviation of the forecast error for simple moving averages in terms of
the standard deviation of each observation.
Having derived the mean and the variance of the forecast error, we still need to specify
the form of the forecast error distribution. By assumption, the values of Dt form a sequence
of independent, identically distributed, normal random variables. Since Ft is a linear combination of Dt�1, Dt�2, . . . . , Dt�N, it follows that Ft is normally distributed and independent of Dt. It now follows that et is normal as well. Hence, the distribution of et is completely
specified by its mean and variance.
As the expected value of the forecast error is zero, we say the method is unbiased.
Notice that this is a result of the assumption that the demand process is stationary. Consider the variance of the forecast error. The value of N that minimizes e is N � � q. This
means that the variance is minimized if the forecast is the average of all the past
data. However, our intuition tells us that we can do better if we use more recent data to
make our forecast. The discrepancy arises because we really do not believe our assumption that the demand process is stationary for all time. A smaller value of N
will allow the moving-average method to react more quickly to unforeseen changes in the
demand process.
nah77856_ch02.indd 123
17/12/14 6:48 PM
nah77856_ch02.qxd
1/17/08
3:15 PM
nah77856
Page 120
124 Chapter Two Forecasting
CASE 2. EXPONENTIAL SMOOTHING
Now consider the case in which forecasts are generated by exponential smoothing. In this
case Ft may be represented by the weighted infinite sum of past values of demand.
Ft � Dt�1 � (1 � )Dt�2 � (1 � )2Dt�3 � � � � ,
E(Ft) � [ � (1 � ) � (1 � )2 � � � �] � .
Notice that this means that E(et) � 0, so that both exponential smoothing and moving
averages are unbiased forecasting methods when the underlying demand process is a constant plus a random term.
Var(Ft) � 22 � (1 � )222 � � � �
� 22 a (1 � )2n.
q
It can be shown that
so that
n�0
1
2n
a (1 � ) � 1 � (1 � )2
n�0
q
Var(Ft) �
Since
22
2
�
.
2
2�
1 � (1 � )
Var(et) � Var(Ft) � Var(Dt),
Var(et) � 2[(2 � ) � 1] � 2[2(2 � )],
or
e � 2
.
A2 � This is the standard deviation of the forecast error for simple exponential smoothing in
terms of the standard deviation of each observation. The distribution of the forecast error
for exponential smoothing is normal for essentially the same reasons as stated above for
moving averages.
Notice that if we equate the variances of the forecast error for exponential smoothing
and moving averages, we obtain
2(2 � ) � (N � 1)N
or � 2(N � 1), which is exactly the same result as we obtained by equating the average
age of data for the two methods.
Appendix 2–B
Derivation of the Equations for the Slope
and Intercept for Regression Analysis
In this appendix we derive the equations for the optimal values of a and b for the regression
model. Assume that the data are (x1, y1), (x2, y2), . . . , (xn, yn), and the regression model to be
fitted is Y � a � bX. Define
g(a, b) � a [yi � (a � bxi)]2.
n
i�1
nah77856_ch02.indd 124
17/12/14 6:48 PM
nah77856_ch02.qxd
1/17/08
3:15 PM
Page 121
Appendix 2–B 125
Interpret g(a, b) as the sum of the squares of the distances from the line a � bx to the
data points yi. The object of the analysis is to choose a and b to minimize g(a, b). This is
accomplished where
0g 0g
� � 0.
0a 0b
That is,
n
0g
� � a 2[yi � (a � bxi)] � 0,
0a
i�1
n
0g
� � a 2xi[yi � (a � bxi)] � 0,
0b
i�1
which results in the two equations
an � b a xi � a yi,
n
n
n
i�1
n
i�1
n
i�1
i�1
i�1
(1)
a a xi � b a x2i � a xi yi.
(2)
These are two linear equations in the unknowns a and b. Multiplying Equation (1) by
� xi and Equation (2) by n gives
an a xi � b a xi
n
i�1
n
i�1
2
� a xi
n
i�1
a y ,
n
(3)
i
i�1
an a xi � bn a x 2i � n a xi yi.
n
n
n
i�1
i�1
i�1
(4)
Subtracting Equation (3) from Equation (4) results in
b n a x 2i � a xi
n
i�1
n
i�1
2
� n a xiyi � a xi
n
i�1
n
i�1
a y .
n
i�1
i
(5)
Define Sxy � n � xiyi � (� xi)(� yi) and Sxx � n� x 2i � (�xi)2. It follows that Equation (5)
may be written bSxx � Sxy, which gives
b�
From Equation (1) we have
or
Sxy
Sxx
.
(6)
an � a yi � b a xi
n
n
i�1
i�1
a � y � bx,
(7)
where y � (1n)� yi and x � (1n)� xi.
These formulas can be specialized to the forecasting problem when the independent variable is assumed to be time. In that case, the data are of the form (1, D1), (2, D2), . . . , (n, Dn),
and the forecasting equation is of the form D̂t � a � bt. The various formulas can be simplified as follows:
n(n � 1)
,
a xi � 1 � 2 � 3 � � � � � n �
2
n(n � 1)(2n � 1)
2
2
.
a xi � 1 � 4 � 9 � � � � � n �
6
nah77856_ch02.indd 125
17/12/14 6:48 PM
nah77856_ch02.qxd
1/17/08
3:15 PM
Page 122
126 Chapter Two Forecasting
Hence, we can write
Sxy � n a iDi � n(n � 1)2 a Di,
n
n
i�1
2
Sxx �
b�
i�1
2
n (n � 1)(2n � 1) n (n � 1)2
�
.
6
4
Sxy
Sxx
b(n � 1)
a�D
� ��
2
Appendix 2–C
Glossary of Notation for Chapter 2
a � Estimate of the intercept in regression analysis.
� Smoothing constant used for single exponential smoothing. One of the smoothing constants used for Holt’s method or one of the smoothing constants used for
Winters’s method.
b � Estimate of the slope in regression analysis.
� Second smoothing constant used for either Holt’s method or Winters’s method.
ct � Seasonal factor for the tth period of a season.
� Third smoothing constant used for Winters’s method.
Dt � Demand in period t. Refers to the series whose values are to be forecasted.
et � Ft � Dt � (Observed) forecasting error in period t.
t � Random variable representing the random component of the demand.
Ft � One-step-ahead forecast made in period t � 1 for demand in period t.
Ft,t� � -step-ahead forecast made in period t for the demand in period t � .
Gt � Smoothed value of the slope for Holt’s and Winters’s methods.
� Mean of the demand process.
MAD � Mean absolute deviation � (1n) g i�1 ei.
n
MAPE � Mean absolute percentage error � (1n) g i�1 eiDi � 100.
MSE � Mean squared error � (1n) g i�1 e2i .
n
n
St � Smoothed value of the series (intercept) for Holt’s and Winters’s methods.
2 � Variance of the demand process.
Tt � Value of the tracking signal in period t (refer to Problem 57).
Bibliography
Armstrong, J. S. “Forecasting by Extrapolation: Conclusions
from Twenty-Five Years of Research.” Interfaces 14 (1984),
pp. 52–66.
Box, G. E. P., and G. M. Jenkins. Time Series Analysis, Forecasting, and Control. San Francisco: Holden Day, 1970.
nah77856_ch02.indd 126
Brown, R. G. Statistical Forecasting for Inventory Control.
New York: McGraw-Hill, 1959.
Brown, R. G. Smoothing, Forecasting, and Prediction of
Discrete Time Series. Englewood Cliffs, NJ: Prentice Hall,
1962.
17/12/14 6:48 PM
Bibliography 127
Cohen, D. "Facebook Has 1.15B Monthly Active Users, But
How Many Are Invalid?" AllFacebook (www.allfacebook.com),
accessed September 3, 2013.
Davenport, W. B., and W. L. Root. An Introduction to the
Theory of Random Signals and Noise. New York:
McGraw-Hill, 1958.
Draper, N. R., and H. Smith. Applied Regression Analysis.
New York: John Wiley & Sons, 1968.
Fisher, M. L.; J. H. Hammond; W. R. Obermeyer;
and A. Raman, “Making Supply Meet Demand in an
­Uncertain World.” Harvard Business Review, May–June
1994, pp. 221–40.
Helmer, O., and N. Rescher. “On the Epistemology of the Inexact Sciences.” Management Science 6 (1959), pp. 25–52.
Kalman, R. E. “A New Approach to Linear Filtering and
Prediction Problems.” Journal of Basic Engineering,
Ser. D 82, 1960, pp. 35–44.
Makridakis, S.; S. C. Wheelwright; and R. J. Hyndman.
Forecasting: Methods and Applications. 3rd ed. New York:
John Wiley & Sons, 1998.
Makridakis, S., and R. L. Winkler. “Averages of Forecasts.”
Management Science 29 (1983), pp. 987–96.
nah77856_ch02.indd 127
McWilliams, G. “At Compaq, a Desktop Crystal Ball.”
Business Week, March 20, 1995, pp. 96–97.
Nahmias, S. “Demand Estimation in Lost Sales Inventory
Systems.” Naval Research Logistics 41 (1994), pp. 739–57.
Nelson, C. R. “The Prediction Performance of the
­FRB-MIT-PENN Model of the U.S. Economy.” The
American Economic Review 62, no. 5 (December 1972),
pp. 902–917.
Nelson, C. R. Applied Time Series Analysis for Managerial
Forecasting. San Francisco: Holden Day, 1973.
Trigg, D. W. “Monitoring a Forecasting System.”
­Operational Research Quarterly 15 (1964), pp. 271–74.
Trigg, D. W., and A. G. Leach. “Exponential Smoothing
with Adaptive Response Rate.” ­Operational Research
­Quarterly 18 (1967), pp. 53–59.
Wiener, N. Extrapolation, Interpolation, and Smoothing
of Stationary Time Series. Cambridge, MA: MIT Press,
1949.
Yule, G. U. “Why Do We Sometimes Get Nonsense
­Correlations between Time Series? A Study of Sampling
and the Nature of Time Series.” Journal of the Royal
­Statistical ­Society 89 (1926), pp. 1–64.
17/12/14 6:48 PM
Chapter Three
Sales and Operations
Planning
“In preparing for battle I have always found that plans are useless, but planning
is indispensable.”
—Dwight D. Eisenhower
Chapter Overview
Purpose
To present the process by which companies go from technical forecasts to aggregate
level sales and operations plans.
Key Points
1. The sales and operations planning (S&OP) process. This chapter could also be
called Macro Planning, since the purpose of an S&OP process is to develop topdown sales and operations plans for the entire firm, or for some subset of the
firm such as a product line or particular plant. The key goals of the process are to
(a) make aggregate level plans that all divisions as well as suppliers can work to;
(b) resolve the inherent tensions between sales and operations divisions; and (c)
anticipate and escalate strategic challenges in matching supply with demand for
the firm.
2. Key Performance Indicators. Inherent in any S&OP process are a set of metrics, or
key performance indicators (KPIs), that the firm uses to judge the performance of
the different divisions. Effective KPIs measure important factors, are relatively easy
to compute, and are actionable, in the sense that those being measured by it can
also effect its change. Operational KPIs may be efficiency or effectiveness focused
and must be aligned with the strategic goals of the firm.
3. The role of uncertainty. It is important to explicitly recognize the role of u
­ ncertainty
in planning. For a firm producing a wide range of products, or providing a
­service, uncertainty can be particularly problematic in planning, and management
­techniques must be adjusted accordingly. Furthermore, different types of ­uncertainty
require different types of management responses. The S&OP process needs to
explicitly acknowledge possible major sources of uncertainty and plan for them
­appropriately.
128
nah77856_ch03.indd 128
17/12/14 6:48 PM
Chapter Three Sales and Operations Planning 129
4. Costs in aggregate operations plans. The following are key costs to be considered
in developing aggregate operations plans.
Smoothing costs. The cost of changing production and/or workforce levels.
Holding costs. The opportunity cost of dollars invested in inventory.
Shortage costs. The costs associated with back-ordered or lost demand.
Labor costs. These include direct labor costs on regular time, overtime,
­subcontracting costs, and idle time costs.
5. Solving aggregate planning problems. Approximate solutions to aggregate
­planning problems can be found graphically and exact solutions via linear
­programming. A level plan has constant production or workforce levels over the
planning horizon, while a chase strategy keeps zero inventory and minimizes
holding and shortage costs. A linear programming formulation assumes that all
costs are linear and typically does not take into account management policy, such
as avoiding hiring and firing as much as possible.
6. Disaggregating plans. While aggregate planning is useful for providing a­ pproximate
solutions for macro planning at the firm level, the question is whether these
­aggregate plans provide any guidance for planning at the lower levels of the firm.
A disaggregation scheme is a mean of taking an aggregate plan and breaking it
down to get more detailed plans at lower levels of the firm.
•
•
•
•
As we go through life, we make both micro and macro decisions. Micro decisions
might be what to eat for breakfast, what route to take to work, what auto service to use,
or which movie to rent. Macro decisions are the kind that change the course of one’s
life: where to live, what to major in, which job to take, whom to marry. A company
also must make both micro and macro decisions every day. Some macro decisions are
highly strategic in nature, such as process technology or market entry choices, and
were discussed in Chapter 1. Other macro decisions are more tactical, such as planning
companywide workforce and production levels or setting sales target. In this chapter
we explore tactical decisions made at the macro level in the context of a process known
as sales and operations planning or S&OP for short.
S&OP begins with demand forecasts and turns them into targets for both sales and
operations plans; techniques for demand forecasting were presented in Chapter 2.
However, a firm will not want to just use such forecasts blindly. They may want to
produce more than is forecast if stock-outs are unacceptable or if they are planning
a major promotion; they may want to produce less than is forecast if overstocks are
costly or they see the product winding down in its lifecycle. Such decisions must be
made at a high strategic level and must involve both the sales and the operations staff.
One of the key goals in S&OP is to resolve the fundamental tension between sales
and operations divisions in an organization. Sales divisions are typically measured
on revenue; they want the product 100% available with as many different varieties as
possible. Meanwhile, operations divisions are typically measured on cost; they want
to keep both capacity and inventory costs low, which means limiting overproduction
and product varieties. The best way to resolve these inherent tensions is with a formal
S&OP process that gets the heads of sales and operations together in a room with other
high level executives. This chapter explores what such processes look like.
Core to the S&OP process is a review of divisional Key Performance Indicators
(KPIs). Section 3.2 reviews key challenges in KPI selection and key types of operational KPIs. One the fundamental challenges in KPI selection is in making sure the
nah77856_ch03.indd 129
17/12/14 6:48 PM
130 Chapter Three Sales and Operations Planning
KPI is both sufficiently high-level to be well aligned with the corporate strategy while
being sufficiently low-level so that it provides a meaningful guide to behavior for
those who are being evaluated by it.
The S&OP process would be a lot simpler if demand were certain; however, as
noted in Chapter 2, it is not and therefore forecasts are generally wrong. In fact, there
are a range of uncertainties that must be considered in the planning process from
known unknowns to unknown unknowns, terms which will be formally defined in
Section 3.3. Part of an S&OP process must include decisions around how to plan for
and mitigate risk or uncertainty.
An important part of S&OP is aggregate planning, which might also be called
macro production planning. It addresses the problem of deciding how many ­employees
the firm should retain and, for a manufacturing firm, the quantity and the mix of
­products to be produced. Macro planning is not limited to manufacturing firms.
­Service o­ rganizations must determine employee staffing needs as well. For example,
airlines must plan staffing levels for flight attendants and pilots, and hospitals must
plan staffing levels for nurses. Macro planning strategies are a fundamental part of the
firm’s overall business strategy. Some firms operate on the philosophy that costs can
be controlled only by making frequent changes in the size and/or composition of the
workforce. The aerospace industry in California in the 1970s adopted this strategy.
As government contracts shifted from one producer to another, so did the technical
workforce. Other firms have a reputation for retaining employees, even in bad times.
Traditionally, IBM and AT&T were two well-known examples.
Aggregate planning methodology is designed to translate demand forecasts into
a blueprint for planning staffing and production levels for the firm over a predetermined planning horizon. Aggregate planning methodology is not limited to top-level
planning. Although generally considered to be a macro planning tool for determining
overall workforce and production levels, large companies may find aggregate planning
useful at the plant level as well. Production planning may be viewed as a hierarchical
process in which purchasing, production, and staffing decisions must be made at several levels in the firm. Aggregate planning methods may be applied at almost any level,
although the concept is one of managing groups of items rather than single items.
This chapter outlines the S&OP process. Core to the process are KPIs, thus this
chapter briefly reviews principles involved in setting KPIs. The chapter also describes
the types of uncertainty that must be planned for in the process and how they are best
addressed. Core to the operations component of any S&OP process is determining an
aggregate production plan for capacity and inventory. This chapter reviews several
techniques for determining aggregate production plans. Some of these are heuristic
(i.e., approximate) and some are optimal. We hope to convey to the reader an understanding of the issues involved in S&OP, a knowledge of the basic tools available for
providing production plans, and an appreciation of the difficulties associated with such
planning in the real world.
3.1 The S&OP Process
The S&OP process is designed to produce a plan that all divisions within in the
­organization, as well as suppliers to the organization, can work to. The process is also
sometimes referred to as sales, inventory, and operations planning (SIOP) to emphasize the important role that inventory can play as a buffer between sales planning and
operations planning.
nah77856_ch03.indd 130
17/12/14 6:48 PM
3.1 The S&OP Process 131
In any organization, demand is a function of sales effort and pricing and supply is
a function of operations effort and capacity. Therefore, in order to best balance supply with demand a strategic approach must be applied. As defined by Grimson and
Pyke (2007), “S&OP is a business process that links the corporate strategic plan to
daily operations plans and enables companies to balance demand and supply for their
­products.”
One of the key words in Grimson and Pike’s definition of S&OP is that it is a
­“business process.” That is, it a consistent set of steps that a company follows on
a ­regular basis. Each business will apply its process slightly differently but some
­common elements include the following:
1. Strategically focused. As discussed above, finding the appropriate balance for supply and demand requires knowledge of the company’s strategic plan. Typically c­ -level
executives are involved in S&OP planning to ensure that the company’s strategy is
­appropriately represented.
2. Cross-functional team. Because the S&OP process must balance the interests of
different parts of the organization, the team performing the process must be both crossfunctional and balanced.
3. Aggregated. It is typically not possible to forecast at the individual stock keeping unit (SKU) level. Thus S&OP processes typically work with aggregated units of
production, such as product families. One natural unit of aggregation is sales dollars,
but of course there is not a one-to-one mapping from sales dollars to units produced.
4. Time fences. The agreed sales and operations plan will typically be only fixed for
a relatively short period (e.g., a week or month) and forecasts for future time ­periods
will typically be assumed to be flexible within a given range. Oftentimes fences are
used to show when the level of certainty changes and can be referred to as frozen
(i.e., fixed), slushy (somewhat fixed), and liquid (highly flexible) time periods.
There are a number of key inputs required for the S&OP process as follows:
1. Technical demand forecast. Advanced forecasting techniques such as those discussed in Chapter 2 are used to understand raw demand.
2. Sales plans. The Marketing and Sales divisions will use the technical forecasts
and their own promotion and marketing plans to provide a sales forecast.
3. Operations plans. Operations will produce a production plan that includes plans
for capacity, supply, and inventory.
4. Innovation plans. The planned role out of new products and future product
­innovation must be considered in the context of sales and operations plans.
5. Financial plans. Any public company must announce earnings forecasts, typically
with a breakdown of how such earnings are to be achieved. Because negative market
surprises can affect a company’s ability to raise capital, the sales and o­ perations plans
must also consider the financial forecasts. Further, in many smaller companies cash flow
constraints must also be considered in the planning process.
The basic S&OP process must iterate between these five key inputs above to resolve
tensions and unforeseen constraints (e.g., a supplier who cannot deliver until six weeks
from today) to finalize the sales and operations plans for the company. It is then left to
the sales and operations divisions of the company to execute on these plans, such as
disaggregating sales dollars into actual production units.
nah77856_ch03.indd 131
17/12/14 6:48 PM
132 Chapter Three Sales and Operations Planning
Figure 3–1
S&OP Overview: Source: Smits and English (2010)
Monthly S&OP Cycle
Weekly Planning Cycle
Draft S&OP Plan:
Demand, Supply
and Inventory Plans
Step 2: Supply
Review
Committed
Demand Plan
Unresolved
Issues
Gap Analysis
Resources
Step 1: Demand
Review
Step 3:
Alignment
Meeting
KPI Dashboards
Demand &
Supply Meeting
Scenarios
Revised marketing
programs or changes to
any other demand
drivers
Committed S&OP
Plan
Unresolved Issues
and Scenarios for
resolution
Proposed S&OP
Plan
Step 4: S&OP
Mgt Review
Meeting
One sample iterative process is shown in Figure 3–1, which has been produced for
Heineken by Smits and English (2010). Notice how the outcome of the S&OP meeting
is a committed S&OP plan for the following month. Key uncertainties are shown as
“scenarios” in the above.
Typically the chief executive officer (CEO) leads the S&OP meeting and makes
the key decisions with respect to trade-offs (Sheldon, 2006). A well run m
­ eeting
is focused on decision making, rather than information dissemination, and all
­participants are expected to have studied the materials that are prepared ahead of
time by the different divisions. The vice president of sales and/or marketing will
be ­present and is the process owner for the demand plan. The vice president of
­operations and/or supply chain management will also be present and is the process
owner for the o­ perations plan. The chief financial officer (CFO) will attend and
prepares the f­inancial plan, which measures performance for the master business
plan. The master scheduler prepares documents and is the process owner for weekly
schedules (which are not typically part of the S&OP process). In a larger organization this will take place at the divisional level, although the CEO may well attend at
least semi-regularly divisional S&OP meetings.
A standard agenda for a monthly S&OP meeting is the following:
• Review of last 30 days—Accuracy of financial, demand, and operations plans
­(typically by product family)
• Review of 30- to 60-day plan expectations—Risks for each plan
nah77856_ch03.indd 132
17/12/14 6:48 PM
Snapshot Application
Heineken International was founded in 1864 by
Gerard Adriaan Heineken in A
­ msterdam and has a
global network of distributors and 125 breweries in
more than 70 countries. Heineken’s corporate strategy is “to be a leading brewer in each of the markets
in which it operates and to have the world’s most
prominent brand portfolio (with Heineken as leading premium beer).” Supply chain planning within
the beer supply chain is complicated by the fact
that most beers, including Heineken, are produced
in batches. In fact, Heineken has a highly prescribed
process for its brewing that all producers must follow to ensure global consistency in taste. It also
has a robust S&OP process, which it implements in
all of its subsidiaries. The key mantra for this process, ­pictured in Figure 3–1, is “one single plan.” This
­process is d
­ escribed as follows:
The global beer market is a dynamic environment
with changing markets, strong competition, and
changing customer preferences. Heineken realized
that a Global Sales and Operations Planning (S&OP)
3.1 The S&OP Process 133
program would become a key enabler in ­supporting
aggressive ­
expansion targets, and would become
necessary to support a retail globalization landscape,
which is applying ­increasing pressures on costs and
service. Heineken’s S&OP process integrates finance,
marketing, sales, and supply chain departments with
the objective of a
­ ligning the organization t­ owards a
synchronized operating plan globally. This p
­ rogram
is supported by a very strong ­project management
­approach which has been designed to ­provide enough
consistency across regions, yet p
­ rovide enough flexibility to embrace and benefit from local cultural
­differences. (Smits & English, 2010)
Notice how this description matches with the ­general
description of S&OP given above yet leaves some flexibility in terms of the actual running of ­meeting to allow
for local customs to apply. The b
­ enefits Heineken has
realized from their S&OP p
­ rocess are better cross-functional decisions; an ­enabler of growth; higher capacity
utilization; lower supply chain costs; and reduced working capital (Rooijen, 2011).
• Review 90- to 120-day risks as required
• Review balance of 12-month horizon for exceptions (Sheldon, 2006)
Typically, an S&OP plan is considered fixed (with agreed windows of flexibility for
later time periods) and, therefore, the operations division must plan for uncertainty
within its own execution plan. Common strategies for dealing with uncertainty include
the following:
1. Buffering: Maintaining excess resources (inventory, capacity, or time) to
cover for fluctuations in supply or demand. Such buffering is an explicit component of planning under uncertainty, such as the inventory models considered in
Chapter 5.
2. Pooling: Sharing buffers to cover multiple sources of variability (e.g., demand
from different markets). This will be covered in more detail in Section 6.7.
3. Contingency planning: Establishing a preset course of action for an anticipated
scenario. This may involve strategies such as bringing on a backup supplier or ­sourcing
from the spot market. This is best done explicitly in the S&OP process.
Heineken Netherlands positions the S&OP process within their planning ­framework
as given in Figure 3–2.
nah77856_ch03.indd 133
17/12/14 6:48 PM
134 Chapter Three Sales and Operations Planning
Figure 3–2
Planning and ­Control
Framework for
Heineken Netherlands.
Source: Every Angle
(2013)
Strategic
Strategy and
Objectives
Long-term
Forecast
Aggregate
Production
Planning
Resource
Planning
Financial
Planning
Sales and
Operations
Planning
Tactical
Firm Orders
Intermediate
Forecast
Demand
management
Operational
Production
Scheduling
Master
Production
Scheduling
Rough-cut
Capacity
Planning
Materials
Requirements
Planning
Capacity
Requirements
Planning
Shop Floor
Control
Input-output
(I/O) Control
Notice how in Figure 3–2, S&OP straddles the strategy and tactical portions of the
planning framework. Many of the other planning topics, such as the master production
schedule and materials requirements planning, will be covered in Chapter 8.
Problems for Section 3.1
1. Why don’t the head of sales or the head of operations chair the S&OP meeting?
2. Why do you think it is important that the CFO attends the S&OP meeting?
3. Describe the likely key tensions between the sales and operations divisions of a
small manufacturer that mostly produces to the local market. Now describe them
for a global organization that sells in many regions and outsources most of its
production to China.
4. Why is S&OP described as an iterative process?
5. What are the trade-offs between having long fixed or frozen periods versus short
fixed or frozen periods as outputs of the S&OP process?
6. What are the trade-offs between using revenue dollars versus sales quantities as
the units in the aggregate S&OP plan.
nah77856_ch03.indd 134
17/12/14 6:48 PM
3.2 Key Performance Indicators 135
3.2 Key Performance Indicators
One of the quantitative inputs to the S&OP process will be the key performance indicators (KPIs) for the divisions and the organization. The CFO is present at the meeting to
present key financial performance measurements such as net profit and return on investment. However, as discussed in Chapter 1, such financial measures tend to e­ mphasize
short-term goals. This issue is often counteracted by the balanced score card approach,
which provides a more well-rounded performance measurement approach covering a
wider scope of metrics than pure finance measures. The balanced score card approach
was developed by Kaplan and Norton (1992) and represents performance measures
­designed to reflect strategy and communicate a vision to the organization. There are four
standard perspectives for the approach, although many organizations have adapted these
for their own use; these perspectives are (a) customer; (b) internal; (c) innovation and
learning; and (d) financial. Figure 3–3 depicts these perspectives.
Figure 3–3
Perspectives for the
Balanced Scorecard
Approach; Source:
Kaplan and Norton
(1992)
Financial
performance
How do we look
to shareholders?
Internal process
performance
What must we
excel at?
Overall strategic
objectives
Customer
performance
How do customers
see us?
Learning and innovation
performance
Can we continue to
improve and create value?
Operational metrics typically sit within internal process measures. There are two
key types of operational KPIs, namely efficiency related KPIs, which are a measure
of resource utilization (e.g., time, cost, materials, etc.) and effectiveness related KPIs,
which are a measure of how well the process meets or exceeds customer requirements
(e.g., defects, complaints, satisfaction score, etc.). Which of these are most important
will depend on the product-line strategy, with efficiency being the most important for
products that compete on cost and effectiveness being the most important for highmargin products that compete on innovation or fashion dimensions.
Some evaluation criteria for the merits of a KPI (ABB, 2010) are: (a) Importance—
Are you measuring the things that really matter? (b) Ease—Does the measurement
‘flow’ from the activity being monitored? (c) Actionable—Can the metric initiate
nah77856_ch03.indd 135
17/12/14 6:48 PM
136 Chapter Three Sales and Operations Planning
a­ ppropriate actions? The KPI must be well understood by employees if it is to be both
measured accurately and to incentivize behavior correctly. One of the key challenges
in choosing KPIs is in achieving organizational alignment. KPIs must be aligned to
the strategic plan in order to monitor the essential performance results implied by the
vision and mission statements and to monitor current strategic goals or objectives. It is
then recommended that they be cascaded down to the business processes to maintain
alignment. Finally, the KPIs may be assigned to business units once they have been
associated with business processes. Going the opposite direction starting from business units is more likely to result in misaligned KPIs. In general, a high level KPI (e.g.,
monthly profit) is easier to align but if employees don’t feel that they can control the
KPI then they may disengage. This is the reason for the recommendation that KPIs be
“actionable” in the above.
In the book The Goal by Goldratt and Cox (1992) it is stated that “The goal of a firm
is to make money.” Many of the issues described in the book arise from misaligned KPIs
to this goal. Goldratt and Cox (1992) suggest that there are really only three operational
KPIs, namely throughput, inventory, and operational expenses; however, their definitions of these terms are nonstandard. Throughput is defined as “the rate at which money
is generated by the system through sales,” inventory is defined as “all the money that
the system has invested in purchasing things it intends to sell,” and operating expenses
are defined as “all the money that the system spends to turn inventory into throughput.”
The beauty of these metrics is that they indeed align incentives with the goal of making
money, the challenge is that firms have found them difficult to implement in practice and
they have not been widely adopted by the field of managerial accounting.
A final challenge in KPI use is gaming of the KPI by employees. If the KPI definition leaves room for interpretation then those being judged by it will typically interpret
it in their favor. One of the most common operational KPIs is DIFOT, which stands
for ­Delivery In-Full and On-Time. However, the definition of “on-time” needs to be
spelled out carefully—does it relate to when the order is shipped or when it arrives
at the customer? Of course, the customer only cares about the latter, but many firms
measure the former as both easier to measure and more fully under operations control
(i.e., more actionable). Clearly there are trade-offs when choosing between the two
perspectives.
Problems for Section 3.2
7. In some S&OP processes all divisions are measured by the same set of KPIs. What
are the advantages and disadvantages to this approach?
8. Give two efficiency-related KPIs and two effectiveness-related KPIs that you
­believe Wal-Mart is likely to use to evaluate its suppliers? Its in-store managers?
9. What types of organizations are likely to face the largest challenges in achieving
alignment between operational KPIs and strategic priorities?
10. Evaluate the following metrics along the three dimensions of importance, ease,
and actionablity: DIFOT, ROI, machine utilization, and number of customer
­complaints.
11. Do you think efficiency-related KPIs or effectiveness-related KPIs will be easier
for employees to “game” in practice? Explain your answer.
12. Why do you think that the KPIs proposed by Goldratt and Cox (1992) have not
been very widely implemented?
nah77856_ch03.indd 136
17/12/14 6:48 PM
3.3 The Role of Uncertainty 137
3.3 The Role of Uncertainty
One of the key challenges in sales and operations planning is the effective management of uncertainty or risk. In a survey by IBM of 400 Supply Chain Executives (IBM,
2009) risk was identified as the second most important key challenge that comprise
the Chief Supply Chain Officer agenda; it was identified by 60 percent of respondents
as important “to a very great extent” or “to a significant extent” (the challenge that
was ranked highest was supply chain visibility). Because S&OP is the highest level of
tactical planning in the organization it must explicitly acknowledge major sources of
uncertainty and explicitly work to manage and/or mitigate risk.
De Meyer et al. (2002) highlight four different types of uncertainty. While they
particularly focus on project management, the categorization also applies to S&OP
processes and operational risks.
1. Variation. Variation is anything that causes the system to depart from regular,
predictable behavior that may be described by a probability distribution (see Appendix
5–A). Typical sources for variation include variability in customer orders, differential
worker skill levels, and variability in quality or lead times.
2. Foreseen uncertainty. Foreseen uncertainty, or “known unknowns” as they
are often called, are risks that can be foreseen and planned for. They are typically
­described by a probability of occurrence and have a larger impact than variation as
described above. Typical sources include supply breakdowns, design changes, natural
disasters, labor strikes, etc.
3. Unforeseen uncertainty. Unforeseen uncertainty, or “unknown unknowns,” are
risks that could not be predicted ahead of time or that are considered too unlikely to
make contingency plans for. Typical sources include natural disasters of unusual scale
or in regions where they do not typically occur (e.g., an earthquake on the eastern coast
of the United States) or competitor innovations that were not anticipated. These are the
so-called “black swan” events of Taleb (2007).
4. Chaos. Chaos is unforeseen uncertainty where the event not only affects the
­operations but also the fundamental goal of the project or company. For example, in
the wake of the New Zealand city of Christchurch’s 2010 earthquake, a number of
local beverage manufacturers offered to switch to bottling water instead of regular
products. Not only was the earthquake unforeseen because the fault rupture occurred
along a previously unknown east-west fault line that is thought to not have moved for
at least 16,000 years, but the goal for the beverage manufacturers shifted from profit
from beverage sales to humanitarian aid.
Each of the above forms of uncertainty must be dealt with and planned for in different ways.
However, all require a firm to have effective management and communication processes.
Variation is the type of uncertainty most commonly dealt with by operations analysts
and managers and will be described in more detail in Section 5.1a. Because variation is
considered routine and within an operations manager purview to anticipate and plan for,
it is not usually much discussed during the S&OP process. Strategies used by operations
to mitigate variation include holding inventory, deliberately maintaining excess capacity,
or incurring deliberate delays in order fulfillment. Foreseen uncertainty is best dealt with
specific contingency plans, risk mitigation processes, and clear ownership responsibility
for processes. It is foreseen uncertainty that is best dealt with explicitly within the S&OP
process. For example, if it is known that a supplier is currently struggling with labor
­issues, then it should be discussed at the S&OP meeting what plans should be put in place
for the event that the supplier’s workforce goes on strike.
nah77856_ch03.indd 137
17/12/14 6:48 PM
138 Chapter Three Sales and Operations Planning
Unforeseen uncertainty may be mitigated by generic contingency plans and good
leadership while chaos requires strong leadership and crisis management processes.
By definition, neither unforeseen uncertainty nor chaos may be explicitly planned
for in the S&OP process. However, an organization with strong cross-functional ties,
which good S&OP processes foster, is going to manage such occurrences better than a
siloed organization with little cross-functional planning.
Problems for Section 3.3
13. Why does the S&OP process typically not explicitly recognize variation even
though it is a fact of life for both operations and sales divisions?
14. Classify the following risks into variation, foreseen uncertainty, unforeseen uncertainty, and chaos:
a. A hurricane on the U.S. East Coast floods a regional warehouse destroying a
large amount of stock.
b. A machine on the plant floor breaks down for an hour.
c. Bad weather on the weekend causes an increase in demand for umbrellas.
d. A cool summer causes a decrease in demand for air conditioners for that season.
e. The excavation process for a new manufacturing plant in the U.S. Midwest
uncovers an archaeological find of such significance that no building can take
place on that site and a new site for the plant must be found.
f. Competitors to the iPad launch smaller tablet computers before the iPad mini is
ready to launch, thus negatively affecting demand for the iPad.
g. The Second World War caused auto manufacturers to switch to producing
­military vehicles.
h. A drug is found to have dangerous side effects following its launch.
i. The transportation disruptions, including the grounding of all airplanes,
­following the attacks on September 11, 2001 severed many supply chains.
15. List two examples each of variation, foreseen uncertainty, and unforeseen uncertainty that you have personally experienced in your studies.
16. Give an example of chaos, either from your own experience or from others, within
the educational domain.
3.4 Aggregate Planning of Capacity
The operations division is responsible for determining an aggregate plan for capacity
usage throughout the planning horizon. This plan uses the agreed upon set of sales
forecasts that comes out of the S&OP meeting. Such a forecast is expressed in terms
of aggregate production units or dollars. The operations division must then determine
aggregate production quantities and the levels of resources required to achieve these
production goals. In practice, this translates to finding the number of workers that
should be employed and the number of aggregate units to be produced in each of the
planning periods 1, 2, …, T. The objective of such aggregate planning is to balance the
advantages of producing to meet demand as closely as possible against the disturbance
caused by changing the levels of production and/or the workforce levels.
As just noted, aggregate planning methodology requires the assumption that
­demand is known with certainty. This is simultaneously a weakness and a strength
nah77856_ch03.indd 138
17/12/14 6:48 PM
3.4 Aggregate Planning of Capacity 139
of the a­pproach. It is a weakness because it does not provide any buffer against
­unanticipated forecast errors. However, most inventory models that allow for random
demand require that the average demand be constant over time. Aggregate planning
allows the manager to focus on the systematic changes that are generally not present in
models that assume random demand. By assuming deterministic demand, the effects
of ­seasonal fluctuations and business cycles can be incorporated into the planning
function. As discussed in Section 3.1, variation caused by demand uncertainty may be
buffered using inventory, capacity, or customer delays.
Costs in Aggregate Capacity Planning
As with most of the optimization problems considered in production management,
the goal of the analysis is to choose the aggregate plan that minimizes cost. It is
important to identify and measure those specific costs that are affected by the
­planning decision.
1. Smoothing costs. Smoothing costs are those costs that accrue as a result of
changing the production levels from one period to the next. In the aggregate planning context, the most salient smoothing cost is the cost of changing the size of the
workforce. Increasing the size of the workforce requires time and expense to advertise
positions, interview prospective employees, and train new hires. Decreasing the size
of the workforces means that workers must be laid off. Severance pay is thus one cost
of decreasing the size of the workforce. However, there are other costs associated with
firing workers that may be harder to measure.
Firing workers could have far-reaching consequences. Firms that hire and fire
f­ requently develop a poor public image. This could adversely affect sales and discourage potential employees from joining the company. It may adversely affect employee
­morale. Furthermore, workers that are laid off might not simply wait around for business to pick up. Firing workers can have a detrimental effect on the future size of
the labor force if those workers obtain employment in other industries. Finally, most
companies are simply not at liberty to hire and fire at will. Labor agreements restrict
the freedom of management to freely alter workforce levels.
Most of the models that we consider assume that the costs of increasing and decreasing the size of the workforce are linear functions of the number of employees that
are hired or fired. That is, there is a constant dollar amount charged for each employee
hired or fired. The assumption on linearity is probably reasonable up to a point. As the
supply of labor becomes scarce, there may be additional costs required to hire more
workers, and the costs of laying off workers may go up substantially if the number of
workers laid off is too large. A typical cost function for changing the size of the workforce appears in Figure 3–4.
2. Holding costs. Holding costs are the costs that accrue as a result of having capital
tied up in inventory. If the firm can decrease its inventory, the money saved could be
invested elsewhere with a return that will vary with the industry and with the s­ pecific
company. (A more complete discussion of holding costs is deferred to Chapter 4.)
­Holding costs are almost always assumed to be linear in the number of units being
held at a ­particular point in time. We will assume for the purposes of the aggregate
­planning a­ nalysis that the holding cost is expressed in terms of dollars per unit held per
­planning ­period. We also will assume that holding costs are charged against the inventory ­remain­ing on hand at the end of the planning period. This a­ ssumption is made for
nah77856_ch03.indd 139
17/12/14 6:48 PM
nah77856_ch03.qxd
1/8/08
2:57 PM
Page 131
140 Chapter Three Sales and Operations Planning
3.3
Costs in Aggregate Planning 131
FIGURE 3–4
$ Cost
Cost of changing the
size of the workforce
Slope = cH
Slope = cF
Ft = Number of fires
Ht = Number of hires
convenience only. Holding costs could be charged against starting inventory or average
inventory as well.
3. Shortage costs. Holding costs are charged against the aggregate inventory as
long as it is positive. In some situations it may be necessary to incur shortages, which
are represented by a negative level of inventory. Shortages can occur when forecasted
demand exceeds the capacity of the production facility or when demands are higher
than anticipated. For the purposes of aggregate planning, it is generally assumed that
excess demand is backlogged and filled in a future period. In a highly competitive situation, however, it is possible that excess demand is lost and the customer goes elsewhere. This case, which is known as lost sales, is more appropriate in the management
of single items and is more common in a retail than in a manufacturing context.
As with holding costs, shortage costs are generally assumed to be linear. Convex
functions also can accurately describe shortage costs, but linear functions seem to be
the most common. Figure 3–5 shows a typical holding/shortage cost function.
4. Regular time costs. These costs involve the cost of producing one unit of output
during regular working hours. Included in this category are the actual payroll costs of
regular employees working on regular time, the direct and indirect costs of materials,
and other manufacturing expenses. When all production is carried out on regular time,
regular payroll costs become a “sunk cost,” because the number of units produced must
equal the number of units demanded over any planning horizon of sufficient length. If
there is no overtime or worker idle time, regular payroll costs do not have to be
included in the evaluation of different strategies.
5. Overtime and subcontracting costs. Overtime and subcontracting costs are the
costs of production of units not produced on regular time. Overtime refers to production by regular-time employees beyond the normal workday, and subcontracting refers
to the production of items by an outside supplier. Again, it is generally assumed that
both of these costs are linear.
6. Idle time costs. The complete formulation of the aggregate planning problem
also includes a cost for underutilization of the workforce, or idle time. In most contexts,
nah77856_ch03.indd 140
17/12/14 6:48 PM
3.4 Aggregate Planning of Capacity 141
Figure 3–5
$ Cost
Holding and backorder costs
Slope = cP
Slope = ci
Back orders
Positive inventory
Inventory
the idle time cost is zero, as the direct costs of idle time would be taken into a­ ccount
in labor costs and lower production levels. However, idle time could have other consequences for the firm. For example, if the aggregate units are input to another process,
idle time on the line could result in higher costs to the subsequent process. In such
cases, one would explicitly include a positive idle cost.
When planning is done at a relatively high level of the firm, the effects of intangible factors are more pronounced. Any solution to the aggregate planning problem ­obtained from a
cost-based model must be considered carefully in the context of company policy. An ­optimal
solution to a mathematical model might result in a policy that requires frequent ­hiring and
firing of personnel. Such a policy may be infeasible b­ ecause of prior contract agreements,
or undesirable because of the potential negative e­ ffects on the firm’s public image.
A Prototype Problem
We will illustrate the various techniques for solving aggregate planning problems with
the following example.
Example 3.1
nah77856_ch03.indd 141
Densepack is to plan workforce and production levels for the six-month period J­anuary to
June. The firm produces a line of disk drives for mainframe computers that are plug compatible
with several computers produced by major manufacturers. Forecast demands over the next six
months for a particular line of drives produced in the Milpitas, California, plant are 1,280, 640,
900, 1,200, 2,000, and 1,400. There are currently (end of December) 300 workers employed
in the Milpitas plant. Ending inventory in December is expected to be 500 units, and the firm
would like to have 600 units on hand at the end of June.
There are several ways to incorporate the starting and the ending inventory constraints into
the formulation. The most convenient is simply to modify the values of the predicted demand.
Define net predicted demand in period 1 as the predicted demand minus initial inventory. If there
is a minimum ending inventory constraint, then this amount should be added to the d
­ emand in
period T. Minimum buffer inventories also can be handled by modifying the predicted demand.
If there is a minimum buffer inventory in every period, this amount should be added to the first
period’s demand. If there is a minimum buffer inventory in only one period, this amount should
be added to that period’s demand and subtracted from the next period’s demand. Actual ending
inventories should be computed using the original demand pattern, however.
17/12/14 6:48 PM
142 Chapter Three Sales and Operations Planning
Returning to our example, we define the net predicted demand for January as 780
(1,280 2 500) and the net predicted demand for June as 2,000 (1,400 1 600). By considering
net demand, we may make the simplifying assumption that starting and ending inventories
are both zero. The net predicted demand and the net cumulative demand for the six months
January to June are as follows:
Net PredictedNet Cumulative
Month
Demand
Demand
January
February
March
April
May
June
780
640
900
1,200
2,000
2,000
780
1,420
2,320
3,520
5,520
7,520
The cumulative net demand is pictured in Figure 3–6. A production plan is the specification of the production levels for each month. If shortages are not permitted, then cumulative
­production must be at least as great as cumulative demand each period. In addition to the
­cumulative net demand, Figure 3–6 also shows one feasible production plan.
In order to illustrate the cost trade-offs of various production plans, we will assume in the
example that there are only three costs to be considered: cost of hiring workers, cost of firing
workers, and cost of holding inventory. Define
cH 5 Cost of hiring one worker 5 $500,
cF 5 Cost of firing one worker 5 $1,000,
cI 5 Cost of holding one unit of inventory for one month 5 $80.
We require a means of translating aggregate production in units to workforce levels. ­Because
not all months have an equal number of working days, we will use a day as an i­ndivisible unit
of measure and define
K 5 Number of aggregate units produced by one worker in one day.
Figure 3–6
8
Cumulative number of units (thousands)
A feasible aggregate
plan for Densepack
7
6
4
Cumulative
net
demand
3
2
Inventory
1
0
nah77856_ch03.indd 142
Feasible
production
schedule
5
0
1
2
3
Month
4
5
6
17/12/14 6:48 PM
3.4 Aggregate Planning of Capacity 143
In the past, the plant manager observed that over 22 working days, with the workforce level
constant at 76 workers, the firm produced 245 disk drives. That means that on average the
production rate was 245y22 5 11.1364 drives per day when there were 76 workers employed
at the plant. It follows that one worker produced an average of 11.1364y76 5 0.14653 drive in
one day. Hence, K 5 0.14653 for this example.
The final data needed to evaluate a plan for this example is the number of working days per
month. In what follows we assume this to be 20, 24, 18, 26, 22, and 15 for January to June,
respectively.
Chase, Level, and Mixed Strategies
Two extremes in capacity planning are the zero inventory plan, also known as a chase
strategy, and the constant workforce plan, also known as a level strategy. Under
the zero inventory plan the workforce is changed each month in order to produce
enough units to most closely match the demand pattern. Capacity is ­adjusted up and
down (i.e., workers are hired and fired) to achieve this matching. Under the constant
­workforce plan, capacity is kept constant during the planning period (i.e., no workers
are hired or fired) and instead inventory is kept between p­ eriods; capacity is set to the
minimum possible to ensure no shortages in any period.
For Example 3–1, a zero inventory plan hires at total of 755 workers, fires a total 145
workers, and achieves a total cost of $572,900 (calculations are not shown and are left as
a reader’s exercise). The best constant workforce plan sets capacity to 411 workers each
month (hiring 111 at the beginning of January) with no further hiring or firing, has a total
inventory cost of $524,960, and a total cost of $580,460 once the initial workers are considered (calculations again left to the reader). While the zero ­inventory plan has slightly
lower costs it is unlikely to be practical because there may be constraints on the total capacity of the plant or on the maximum change that is p­ ossible from one month to the next.
The zero inventory plan and constant workforce strategies are pure strategies: they
are designed to achieve one objective. They are useful in enhancing intuition and for
ballpark calculations. However, with more flexibility, small ­modifications can result
in dramatically lower costs. Such plans are described as mixed ­strategies. Optimal
mixed strategies are typically found using linear programming formulations, which
can ­incorporate a variety of additional practical constraints. Linear programming
­formulations are the subject of the next section.
Problems for Sections 3.4
17. A local machine shop employs 60 workers who have a variety of skills. The shop
­accepts one-time orders and also maintains a number of regular clients. Discuss some
of the difficulties with using the aggregate planning methodology in this c­ ontext.
18. A large manufacturer of household consumer goods is considering integrating an
aggregate planning model into its manufacturing strategy. Two of the company vice
presidents disagree strongly as to the value of the approach. What arguments might
each of the vice presidents use to support his or her point of view?
19. Describe the following costs and discuss the difficulties that arise in attempting to
measure them in a real operating environment.
a. Smoothing costs
b. Holding costs
c. Payroll costs
nah77856_ch03.indd 143
17/12/14 6:48 PM
Snapshot Application
144 Chapter Three Sales and Operations Planning
HP Enterprise Services Uses Optimization
for Workforce Planning
The S&OP planning problem discussed in this chapter
includes optimization of several of the firm’s functions,
including production planning and workforce planning. In practice, the workforce planning problem is
much more difficult than the problem framed in this
chapter, since it is often the case that both the supply and demand for workers is uncertain. This means
that both demand and supply are likely to be random
variables, thus making the problem potentially very
complex. HP Enterprise Services (HPES) is a part of the
HP Corporation that focuses on global business and
technology. These services are provided by more than
100,000 employees located in over 60 countries. To
provide an optimal solution to HPES’ planning problem
is probably not realistic due to the scale of the problem. Traditional operations research methods such as
Markov Decision Process modeling or large scale mixed
integer programming (two methods that have been
proposed in the literature) quickly become computationally unwieldy.
A group of researchers (Santos, et. al. (2013)) suggested a two stage approach for solving this problem.
The first stage (supply and demand consolidation) indicates those jobs for which an employee is fully qualified.
For those employees that are partially qualified, they
develop a transition table that provides scores which
indicate the degree of qualification. Supply uncertainty
in this context is primarily due to employee attrition,
which can be 30% or more in some locations. Under
ideal circumstances, the firm would like to be able to
match job requirements with those employees that are
100% qualified. However, such a stringent rule results
in poor demand fulfilment levels. Rather, the group
designed a flexible matching scheme based c­oncepts
developed for the analytical hierarchy process, a tool
that allows one to determine appropriate weights in a
multi-criteria decision problem.
The second stage of the analysis is to build a mixed
integer programming (MIP) module to allocate ­workers
to jobs. This is done in stages: first, available employees are allocated to jobs. Second, employees that are
currently committed to jobs, but will be freed up at
some future time are allocated. If neither of these
schemes covers the job requirements, then new hires
are recommended. Finally, if positions are still unfilled,
a “gap” is declared. As noted above, mixed integer
optimization can be very demanding computationally,
so the authors employ a heuristic for this phase. Gaps
are quite common, so procedures are ­recommended to
deal with them.
HP first implemented this approach at its facility in
Bangalore, India in 2009. Resource utilization rates improved from approximately 75% to approximately 90%
as a result of this planning tool. HP is now in the process
of implementing this system on a worldwide scale.
Source: Santos, C., et al. “HP Enterprise Services Uses
­Optimization for Resource Planning”. Interfaces 43 (2013),
­pp. 152–169.
20. Discuss the following statement: “Since we use a rolling production schedule, I
­really don’t need to know the demand beyond next month.”
21. St. Clair County Hospital is attempting to assess its needs for nurses over the
coming four months (January to April). The need for nurses depends on both the
­numbers and the types of patients in the hospital. Based on a study conducted
by consultants, the hospital has determined that the following ratios of nurses to
­patients are ­required:
Numbers of Nurses
Patient TypeRequired per Patient
Major surgery
Minor surgery
Maternity
Critical care
Other
nah77856_ch03.indd 144
0.4
0.1
0.5
0.6
0.3
Patient Forecasts
Jan.
Feb.
28
12
22
75
80
21
25
43
45
94
Mar.Apr.
16
45
90
60
73
18
32
26
30
77
17/12/14 6:48 PM
3.5 Solving Aggregate Planning Problems 145
a. How many nurses should be working each month to most closely match ­patient
forecasts?
b. Suppose the hospital does not want to change its policy of not increasing the
nursing staff size by more than 10 percent in any month. Suggest a schedule of
nurse staffing over the four months that meets this requirement and also meets
the need for nurses each month.
22. Give an example of an environment where a chase strategy would be highly
­disruptive. Now give one where a level strategy would be highly disruptive.
23. Is a chase or level strategy more appropriate for aggregate planning in an air conditioning manufacturing plant where demand is highly seasonal and the workforce
is relatively skilled? Explain your answer.
3.5 Solving Aggregate Planning Problems
Linear programming is a term used to describe a general class of optimization problems.
The objective is to determine values of n nonnegative real variables in order to maximize or minimize a linear function of these variables that is subject to m linear constraints of these variables.1 The primary advantage in formulating a problem as a linear
program is that optimal solutions can be found very efficiently by the simplex method.
When all cost functions are linear, there is a linear programming formulation of the
general aggregate planning problem. Because of the efficiency of commercial linear
programming codes, this means that (essentially) optimal solutions can be obtained for
very large problems.2
Cost Parameters and Given Information
The following values are assumed to be known:
cH 5 Cost of hiring one worker,
cF 5 Cost of firing one worker,
cI 5 Cost of holding one unit of stock for one period,
cR 5 Cost of producing one unit on regular time,
cO 5 Incremental cost of producing one unit on overtime,
cU 5 Idle cost per unit of production,
cS 5 Cost to subcontract one unit of production,
nt 5 Number of production days in period t,
K 5 Number of aggregate units produced by one worker in one day,
I0 5 Initial inventory on hand at the start of the planning horizon,
W0 5 Initial workforce at the start of the planning horizon,
Dt 5 Forecast of demand in period t.
The cost parameters also may be time dependent; that is, they may change with t.
­ ime-dependent cost parameters could be useful for modeling changes in the costs of hiring
T
or firing due, for example, to shortages in the labor pool, or changes in the costs of production and/or storage due to shortages in the supply of resources, or changes in i­ nterest rates.
1
An overview of linear programming can be found in Supplement 1, which follows this chapter.
The qualifier is included because rounding may give suboptimal solutions. There will be more about this
point later.
2
nah77856_ch03.indd 145
17/12/14 6:48 PM
146 Chapter Three Sales and Operations Planning
Problem Variables
The following are the problem variables:
Wt 5 Workforce level in period t,
Pt 5 Production level in period t,
It 5 Inventory level in period t,
Ht 5 Number of workers hired in period t,
Ft 5 Number of workers fired in period t,
Ot 5 Overtime production in units,
Ut 5 Worker idle time in units (“undertime”),
St 5 Number of units subcontracted from outside.
The overtime and idle time variables are determined in the following way. The term Knt
represents the number of units produced by one worker in period t, so that KntWt would be
the number of units produced by the entire workforce in period t. However, we do not require that KntWt 5 Pt. If Pt . KntWt, then the number of units produced ­exceeds what the
workforce can produce on regular time. This means that the difference is being produced
on overtime, so that the number of units produced on overtime is exactly Ot 5 Pt 2 KntWt.
If Pt , KntWt, then the workforce is pro­ducing less than it should be on regular time,
which means that there is worker idle time. The idle time is measured in units of production rather than in time, and is given by Ut 5 KntWt 2 Pt.
Problem Constraints
Three sets of constraints are required for the linear programming formulation. They
are included to ensure that conservation of labor and conservation of units are satisfied.
1. Conservation of workforce constraints.
Wt
5 Wt21 1 Ht 2 Ft
Number 5 Number 1 Number 2 Number
of workers of workers hired
fired
in tin t 2 1 in t
in t
for 1 # t # T.
2. Conservation of units constraints.
5 It21 1
Pt
1
St
2 Dt
It
Inventory 5 Inventory 1 Number 1 Number 2 Demand
in t
in t 2 1
of units
of units
in t
produced subcontracted
in t
in t
for 1 # t # T.
3. Constraints relating production levels to workforce levels.
5 KntWt 1
Ot
2
Ut
Pt
Number 5 Number 1 Number 2 Number
of units
of units
of units
of units
produced
produced produced
of idle
in t
by regular
on overproduction
workforce
time in t
in t
in t
nah77856_ch03.indd 146
for 1 # t # T.
17/12/14 6:48 PM
nah77856_ch03.qxd
1/8/08
2:57 PM
Page 143
3.5
Solution of Aggregate3.5 Planning
Problems
by Linear
Programming
143
Solving
Aggregate
Planning
Problems 147
In addition to these constraints, linear programming requires that all problem variables
be nonnegative. These constraints and the nonnegativity constraints are the minimum that
must be present in any formulation. Notice that (1), (2), and (3) constitute 3T constraints,
rather than 3 constraints, where T is the length of the forecast horizon.
The formulation also requires specification of the initial inventory, I0, and the initial workforce, W0, and may include specification of the ending inventory in the final
period, IT.
The objective function includes all the costs defined earlier. The linear programming
formulation is to choose values of the problem variables Wt, Pt, It, Ht, Ft, Ot, Ut, and St to
Minimize a (cH Ht � cF Ft � cI It � cR Pt � cO Ot � cU Ut � cSSt)
T
t�1
subject to
Wt � Wt�1 � Ht � Ft
for 1 � t � T
(conservation of workforce),
(A)
Pt � KntWt � Ot � Ut
for 1 � t � T
(production and workforce)
(B)
It � It�1 � Pt � St � Dt
for 1 � t � T (inventory balance),
Ht, Ft, It, Ot, Ut, St, Wt, Pt � 0 (nonnegativity),
(C)
(D)
plus any additional constraints that define the values of starting inventory, starting
workforce, ending inventory, or any other variables with values that are fixed in advance.
Rounding the Variables
In general, the optimal values of the problem variables will not be integers. However,
fractional values for many of the variables do not make sense. These variables include
the size of the workforce, the number of workers hired each period, and the number of
workers fired each period, and also may include the number of units produced each
period. (It is possible that fractional numbers of units could be produced in some
applications.) One way to deal with this problem is to require in advance that some or
all of the problem variables assume only integer values. Unfortunately, this makes the
solution algorithm considerably more complex. The resulting problem, known as an integer linear programming problem, requires much more computational effort to solve
than does ordinary linear programming. For a moderate-sized problem, solving the
problem as an integer linear program is certainly a reasonable alternative.
If an integer programming code is unavailable or if the problem is simply too large
to solve by integer programming, linear programming still provides a workable solution. However, after the linear programming solution is obtained, some of the problem
variables must be rounded to integer values. Simply rounding off each variable to the
closest integer may lead to an infeasible solution and/or one in which production and
workforce levels are inconsistent. It is not obvious what is the best way to round the
variables. We recommend the following conservative approach: round the values of the
numbers of workers in each period t to Wt, the next larger integer. Once the values of
Wt are determined, the values of the other variables, Ht , Ft , and Pt , can be found along
with the cost of the resulting plan.
Conservative rounding will always result in a feasible solution, but will rarely
give the optimal solution. The conservative solution generally can be improved by
trial-and-error experimentation.
nah77856_ch03.indd 147
17/12/14 6:48 PM
nah77856_ch03.qxd
1/8/08
2:57 PM
Page 144
144
Chapter Three
Aggregate
Planning Planning
148 Chapter
Three Sales
and Operations
There is no guarantee that if a problem can be formulated as a linear program, the
final solution makes sense in the context of the problem. In the aggregate planning
problem, it does not make sense that there should be both overtime production and idle
time in the same period, and it does not make sense that workers should be hired and
fired in the same period. This means that either one or both of the variables Ot and Ut
must be zero, and either one or both of the variables Ht and Ft must be zero for each t,
1 � t � T. This requirement can be included explicitly in the problem formulation by
adding the constraints
OtUt � 0
for 1 � t � T,
Ht Ft � 0
for 1 � t � T,
since if the product of two variables is zero it means that at least one must be zero. Unfortunately, these constraints are not linear, as they involve a product of problem
variables. However, it turns out that it is not necessary to explicitly include these
constraints, because the optimal solution to a linear programming problem always occurs at an extreme point of the feasible region. It can be shown that every extreme point
solution automatically has this property. If this were not the case, the linear programming solution would be meaningless.
Extensions
Linear programming also can be used to solve somewhat more general versions of the
aggregate planning problem. Uncertainty of demand can be accounted for indirectly by
assuming that there is a minimum buffer inventory Bt each period. In that case we
would include the constraints
It � Bt
for 1 � t � T.
The constants Bt would have to be specified in advance. Upper bounds on the number of workers hired and the number of workers fired each period could be included in
a similar way. Capacity constraints on the amount of production each period could
easily be represented by the set of constraints:
Pt � Ct
for 1 � t � T.
The linear programming formulation introduced in this section assumed that inventory levels would never go negative. However, in some cases it might be desirable or
even necessary to allow demand to exceed supply, for example, if forecast demand
exceeded production capacity over some set of planning periods. In order to treat
backlogging of excess demand, the inventory level It must be expressed as the difference between two nonnegative variables, say It� and It�, satisfying
It � It� � It�,
It� � 0,
It� � 0.
The holding cost would now be charged against It� and the penalty cost for back orders (say cP) against It�. However, notice that for the solution to be sensible, it must be
true that I t� and It� are not both positive in the same period t. As with the overtime and
idle time and the hiring and firing variables, the properties of linear programming will
guarantee that this holds without having to explicitly include the constraint It�It� � 0
in the formulation.
nah77856_ch03.indd 148
17/12/14 6:48 PM
nah77856_ch03.qxd
1/8/08
2:57 PM
Page 145
Solution of Aggregate3.5 Planning
Problems
by Linear
Programming
145
Solving
Aggregate
Planning
Problems 149
3.5
FIGURE 3–7
A convex piecewiselinear function
In the development of the linear programming model, we stated the requirement that
all the cost functions must be linear. This is not strictly correct. Linear programming
also can be used when the cost functions are convex piecewise-linear functions.
A convex function is one with an increasing slope. A piecewise-linear function is
one that is composed of straight-line segments. Hence, a convex piecewise-linear
function is a function composed of straight lines that have increasing slopes. A typical
example is presented in Figure 3–7.
In practice, it is likely that some or all of the cost functions for aggregate planning
are convex. For example, if Figure 3–7 represents the cost of hiring workers, then the
marginal cost of hiring one additional worker increases with the number of workers
that have already been hired. This is probably more accurate than assuming that the
cost of hiring one additional worker is a constant independent of the number of
workers previously hired. As more workers are hired, the available labor pool shrinks
and more effort must be expended to hire the remaining available workers.
In order to see exactly how convex piecewise-linear functions would be incorporated
into the linear programming formulation, we will consider a very simple case. Suppose
that the cost of hiring new workers is represented by the function pictured in Figure 3–8.
According to the figure, it costs cH1 to hire each worker until H* workers are hired, and
it costs cH2 for each worker hired beyond H* workers, with cH1 cH2. The variable Ht,
the number of workers hired in period t, must be expressed as the sum of two variables:
Ht H1t H2t.
Interpret H1t as the number of workers hired up to H* and H2t as the number of
workers hired beyond H* in period t. The cost of hiring is now represented in the
objective function as
a (cH1H1t cH2 H2t),
T
t1
nah77856_ch03.indd 149
17/12/14 6:48 PM
150 Chapter Three Sales and Operations Planning
Figure 3–8
Convex piecewise­linear hiring cost
function
Slope = cH2
Slope = cH1
H*
and the additional constraints
Ht 5 H1t 1 H2t
0 # H1t # H*
0 # H2t
must also be included.
In order for the final solution to make sense, it can never be the case that H1t , H*
and H2t . 0 for some t. (Why?) However, because linear programming searches for
the minimum cost solution, it will force H1t to its maximum value before allowing H2t
to become positive, since cH1 , cH2. This is the reason that the cost functions must be
convex. This approach can easily be extended to more than two linear segments and
to any of the other cost functions present in the objective function. The technique is
known as separable convex programming and is discussed in greater detail in Hillier
and Lieberman (1990).
We will demonstrate the use of linear programming by finding the optimal solution
to the example presented in Section 3.4. As there is no subcontracting, overtime, or
idle time allowed, and the cost coefficients are constant with respect to time, the objective function is simply
Minimize a 500 a Ht + 1,000 a Ft + 80 a It b .
nah77856_ch03.indd 150
6
6
6
t=1
t=1
t=1
17/12/14 6:48 PM
3.5 Solving Aggregate Planning Problems 151
The boundary conditions comprise the specifications of the initial inventory of
500 units, the initial workforce of 300 workers, and the ending inventory of 600 units.
These are best handled by including a separate additional constraint for each boundary
condition.
The constraints are obtained by substituting t 5 1, . . . , 6 into Equations (A), (B),
and (C). The full set of constraints expressed in standard linear programming format
(with all problem variables on the left-hand side and nonnegative constants on the
right-hand side) is as follows:
W1 2 W0 2 H1 1 F1 5 0,
W2 2 W1 2 H2 1 F2 5 0,
W3 2 W2 2 H3 1 F3 5 0,
W4 2 W3 2 H4 1 F4 5 0,
(A)
W5 2 W4 2 H5 1 F5 5 0,
W6 2 W5 2 H6 1 F6 5 0;
P1 2 I1 1 I0 5 1,280,
P2 2 I2 1 I1 5 640,
P3 2 I3 1 I2 5 900,
P4 2 I4 1 I3 5 1,200,
(B)
P5 2 I5 1 I4 5 2,000,
P6 2 I6 1 I5 5 1,400;
P1 2 2.931W1 5 0,
P2 2 3.517W2 5 0,
P3 2 2.638W3 5 0,
P4 2 3.810W4 5 0,
(C)
P5 2 3.224W5 5 0,
P6 2 2.198W6 5 0;
W1, . . . , W6, P1, . . . , P6, I1, . . . , I6, F1, . . . , F6, H1, . . . , H6 $ 0;
(D)
W0 5 300,
I0 5 500,
(E)
I6 5 600.
nah77856_ch03.indd 151
17/12/14 6:48 PM
152 Chapter Three Sales and Operations Planning
The values in equations (C) come from multiplying the value of K, number of aggregate units per worker,
found earlier by the number of working days in the month, which shows this formulation in Excel.
A
1
2
3
4
5
B
Cost of hiring
C
D
E
F
G
H
I
J
$500
Cost of firing
$1,000
Holding cost
$80
K
0.14653
Ending inv
600
6
7
Month
8
Hired
Fired
Inventory
Workers Production
(H_t)
(F_t)
(I_t)
(W_t)
9
Start
10
January
0.00
(P_t)
Worker Inventory Adjusted
balance balance
Demand
500
300
27.02
20.00
272.98
800.00
0
1280
1280
11
February
0.00
0.00
340.00
272.98
960.00
0
640
640
12
March
0.00
0.00
160.00
272.98
720.00
0
900
900
13
April
0.00
0.00
0.00
272.98
1040.00
0
1200
1200
14
May
464.80
0.00
378.38
737.78
2378.38
0
2000
2000
15
June
0.00
600.00
737.78
1621.62
0
1400
1400
16
Totals
0.00
464.80
27.02 1498.38
17
18
Month
19
Days
Units per Production
worker
(n_t)
balance
20
January
20
2.931
0
21
February
24
3.517
0
22
March
18
2.638
0
23
April
26
3.810
0
24
May
22
3.224
0
25
June
15
2.198
0
26
27
Total cost
$379,292.22
Cell Formulas
Cell FormulaCopied to
B4 =245/22/76
H9 =E9-E8-B9+C9
H10:H14
I9 =F9-D9+D8I10:I14
C20 =B20*$B$4C21:C25
D20 =F10-C20*E10
D21:D25
B16 =SUM(B10:B15)
C16:D16
B27 =$B$1*$B$16+$B$2*$C$16+$B$3*$D$16
nah77856_ch03.indd 152
17/12/14 6:48 PM
3.5 Solving Aggregate Planning Problems 153
The Solver window for this problem is as follows.
nah77856_ch03.indd 153
17/12/14 6:48 PM
154 Chapter Three Sales and Operations Planning
Table 3–1 Aggregate Plan for Densepack Obtained from Rounding the Linear Programming Solution
AB C
DE
FGHI
Number
Number
of UnitsEnding
Number ofNumberNumber
of Units
Produced Cumulative CumulativeInventory
Month
WorkersHired
Fired
per Worker
(B 3 E)
ProductionNet Demand
(G 2 H)
January
273
27
2.931 800 800 780 20
February
273
3.517 960
1,760
1,420
340
March
273
2.638 720
2,480
2,320
160
April
273
3.810
1,040
3,520
3,520 0
May
738
465 3.224
2,379
5,899
5,520
379
June
738
2.198
1,622
7,521
7,520 1
Totals 465
27
900
The value of the objective function at the optimal solution is 379,292.22,which is
considerably less than that achieved with either the zero inventory plan or the constant
workforce plan. However, this cost is based on fractional values of the variables. The
actual cost will be slightly higher after rounding.
Following the rounding procedure recommended earlier, we will round all the
­val­ues of Wt to the next higher integer. That gives W1 5 ? ? ? 5 W4 5 273 and W5 5
W6 5 738. This determines the values of the other problem variables. This means that
the firm should fire 27 workers in January and hire 465 workers in May. The complete
­solution is given in Table 3–6.
Again, because column H in Table 3–1 corresponds to net demand, we add the
600 units of ending inventory in June, giving a total inventory of 900 1 600 5 1,500
units. Hence, the total cost of this plan is (500)(465) 1 (1,000)(27) 1 (80)(1,500) 5
$379,500, which represents a substantial savings over both the zero inventory plan and
the constant workforce plan.
The results of the linear programming analysis suggest another plan that might be
more suitable for the company. Because the optimal strategy is to decrease the workforce in January and build it back up again in May, a reasonable alternative might
be to not fire the 27 workers in January and to hire fewer workers in May. In this
case, the most efficient method for finding the correct number of workers to hire in
May is to simply re-solve the linear program, but without the variables F1, . . . , F6,
as no firing of workers means that these variables are forced to zero. (If you wish to
avoid reentering the problem into the computer, simply append the old formulation
with the constraints F1 5 0, F2 5 0, . . . , F6 5 0.) The optimal number of workers
to hire in May turns out to be 374 if no workers are fired, and the cost of the plan is
approximately $386,120. This is only slightly more expensive than the optimal plan,
and has the important advantage of not requiring the firing of any workers.
Problems for Section 3.5
24. Mr. Meadows Cookie Company makes a variety of chocolate chip cookies in
the plant in Albion, Michigan. Based on orders received and forecasts of buying habits, it is estimated that the demand for the next four months is 850,
nah77856_ch03.indd 154
17/12/14 6:48 PM
3.5 Solving Aggregate Planning Problems 155
1,260, 510, and 980, expressed in thousands of cookies. During a 46-day
­period when there were 120 workers, the company produced 1.7 million cookies. ­Assume that the number of workdays over the four months are respectively
26, 24, 20, and 16. There are c­ urrently 100 workers employed, and there is no
­starting inventory of cookies.
a. What is the minimum constant workforce required to meet demand over the
next four months?
b. Assume that cI 5 10 cents per cookie per month, cH 5 $100, and c F 5 $200.
Evaluate the cost of the plan derived in part (a).
c. Formulate as a linear program. Be sure to define all variables and include the
required constraints.
d. Solve for the optimal solution.
25. Harold Grey owns a small farm in the Salinas Valley that grows apricots. The apricots are dried on the premises and sold to a number of large supermarket chains.
Based on past experience and committed contracts, he estimates that sales over the
next five years in thousands of packages will be as follows:
Year
Forecasted Demand
(thousands of packages)
1
2
3
4
5
300
120
200
110
135
Assume that each worker stays on the job for at least one year, and that Grey
currently has three workers on the payroll. He estimates that he will have 20,000
packages on hand at the end of the current year. Assume that, on the average, each
worker is paid $25,000 per year and is responsible for producing 30,000 packages. Inventory costs have been estimated to be 4 cents per package per year, and
­shortages are not allowed.
Based on the effort of interviewing and training new workers, Farmer Grey
­estimates that it costs $500 for each worker hired. Severance pay amounts to
$1,000 per worker.
a. Assuming that shortages are not allowed, determine the minimum constant
workforce that he will need over the next five years.
b. Evaluate the cost of the plan found in part (a).
c. Formulate this as a linear program.
d. Solve the problem and round-off the solution and determine the cost of the
­resulting plan.
26. A local semiconductor firm, Superchip, is planning its workforce and ­production
levels over the next year. The firm makes a variety of microprocessors and uses
sales dollars as its aggregate production measure. Based on orders received
nah77856_ch03.indd 155
17/12/14 6:48 PM
156 Chapter Three Sales and Operations Planning
and sales forecasts provided by the marketing department, the estimate of dollar
sales for the next year by month is as follows:
Month
January
February
March
April
May
June
July
August
September
October
November
December
Production
Days
Predicted Demand
(in $10,000)
22
16
21
19
23
20
24
12
19
22
20
16
340
380
220
100
490
625
375
310
175
145
120
165
Inventory holding costs are based on a 25 percent annual interest charge. It is ­anticipated
that there will be 675 workers on the payroll at the end of the current year and inventories
will amount to $120,000. The firm would like to have at least $100,000 of inventory at the
end of December next year. It is e­ stimated that each worker accounts for an average of
$60,000 of production per year (assume that one year consists of 250 working days). The
cost of hiring a new worker is $200, and the cost of laying off a worker is $400.
a. Formulate this as a linear program.
b. Solve the problem. Round the variables in the resulting solution and determine
the cost of the plan you obtain.
27.Consider Mr. Meadows Cookie Company, described in Problem 24. Suppose that
the cost of hiring workers each period is $100 for each worker until 20 workers are
hired, $400 for each worker when between 21 and 50 workers are hired, and $700
for each worker hired beyond 50.
a. Write down the complete linear programming formulation of the revised problem.
b. Solve the revised problem for the optimal solution. What difference does the
new hiring cost function make in the solution?
28. Leather-All produces a line of handmade leather products. At the present time,
the company is producing only belts, handbags, and attaché cases. The predicted
demand for these three types of items over a six-month planning h­ orizon is as
­follows:
Number of
Month
Working DaysBeltsHandbagsAttaché Cases
1
2
3
4
5
6
nah77856_ch03.indd 156
22
20
19
24
21
17
2,500
2,800
2,000
3,400
3,000
1,600
1,250
680
1,625
745
835
375
240
380
110
75
126
45
17/12/14 6:48 PM
3.6 Disaggregating Plans 157
The belts require an average of two hours to produce, the handbags three hours, and
the attaché cases six hours. All the workers have the skill to work on any item.
Leather-All has 46 employees who each has a share in the firm and cannot be fired.
There are an additional 30 locals that are available and can be hired for short periods at higher cost. Regular employees earn $8.50 per hour on regular time and
$14.00 per hour on overtime. Regular time comprises a seven-hour workday, and
the regular employees will work as much overtime as is available. The additional
workers are hired for $11.00 per hour and are kept on the payroll for at least one
full month. Costs of hiring and firing are negligible.
Because of the competitive nature of the industry, Leather-All does not want to
incur any demand back orders.
a. Using worker hours as an aggregate measure of production, convert the
­fore­casted demands to demands in terms of aggregate units.
b. What would be the size of the workforce needed to satisfy the demand
for the coming six months on regular time only? Would it be to the company’s ­advantage to bring the permanent workforce up to this level? Why
or why not?
c. Formulate the problem of optimizing Leather-All’s hiring schedule as a l­inear
program. Define all problem variables and include whatever ­constraints are
­necessary.
d. Solve the problem formulated in part (a) for the optimal solution. Round all
the relevant variables and determine the cost of the r­ esulting plan.
3.6 Disaggregating Plans
Aggregate
As we saw earlier in this chapter, aggregate planning may be done at several levels
of the firm. Example 3.1 considered a single plant, and showed how one might define
an aggregate unit to include the six different items produced at that plant. Aggregate
planning might be done for a single plant, a product family, a group of families, or for
the firm as a whole.
There are two views of production planning: bottom-up or top-down. The bottomup approach means that one would start with individual item production plans. These
plans could then be aggregated up the chain of products to produce aggregate plans.
The top-down approach, which is the one treated in this chapter, is to start with an
aggregate plan at a high level. These plans would then have to be “disaggregated” to
produce detailed production plans at the plant and individual item levels.
It is not clear that disaggregation is an issue in all circumstances. If the aggregate
plan is used only for macro planning purposes, and not for planning at the detail level,
then one need not worry about disaggregation. However, if it is important that individual item production plans and aggregate plans be consistent, then it might be necessary
to consider disaggregation schemes.
The disaggregation problem is similar to the classic problem of resource allocation. Consider how resources are allocated in a university, for example. A university receives revenues from tuition, gifts, interest on the endowment, and research
grants. Costs include salaries, maintenance, and capital expenditures. Once an
nah77856_ch03.indd 157
17/12/14 6:48 PM
158 Chapter Three Sales and Operations Planning
a­ nnual budget is determined, each school (arts and science, engineering, business,
law, etc.) and each budget center (staff, maintenance, buildings and grounds, etc.)
would have to be allocated its share. Budget centers would have to allocate funds
to each of their subgroups. For example, each school would allocate funds to individual departments, and departments would allocate funds to faculty and staff in
that department.
In the manufacturing context, a disaggregation scheme is just a means of allocating aggregate units to individual items. Just as funds are allocated on several levels
in the university, aggregate units might have to be disaggregated at several levels
of the firm. This is the idea behind hierarchical production planning championed by
several researchers at MIT and reported in detail in Bitran and Hax (1981) and Hax
and Candea (1984).
We will discuss one possible approach to the disaggregation problem from B
­ itran
and Hax (1981). Suppose that X* represents the number of aggregate units of production for a particular planning period. Further, suppose that X* represents an
­aggregation of n different items (Y1, Y2, . . . , Yn). The question is how to d­ ivide up
(i.e., disaggregate) X* among the n items. We know that holding costs are a­ lready
­included in the determination of X*, so we need not include them again in the
­disaggregation scheme. Suppose that Kj represents the fixed cost of setting up for
­production of Yj, and lj is the annual usage rate for item j. A reasonable optimization
criterion in this context is to choose Y1, Y2, . . . , Yn to minimize the average annual
cost of setting up for production. As we will see in Chapter 4, the average annual
setup cost for item j is K j lj >Y j. Hence, disaggregation requires solving the following
mathematical programming problem:
Minimise a
J
K jlj
J= 1 Y j
subject to
aYj = X*
j=1
J
and
aj # Yj # bj for 1 # j # J.
The upper and lower bounds on Yj account for possible side constraints on the
­production level for item j.
A number of feasibility issues need to be addressed before the family run sizes Yj
are further disaggregated into lots for individual items. The objective is to schedule
the lots for individual items within a family so that they run out at the scheduled setup
time for the family. In this way, items within the same family can be produced within
the same production setup.
nah77856_ch03.indd 158
17/12/14 6:48 PM
nah77856_ch03.qxd
1/24/08
1:46 PM
Page 157
Snapshot Application
159
WELCH’S USES AGGREGATE PLANNING FOR
PRODUCTION SCHEDULING
Welch’s is a make-to-stock food manufacturer based in
Concord, Massachusetts. They are probably best known
for grape jelly and grape juices, but they produce a
wide variety of processed foods. Allen and Schuster
(1994) describe an aggregate planning model for
Welch’s primary production facility.
The characteristics of the production system for
which their system was designed are
• Dynamic, uncertain demand, resulting in changing
buffer stock requirements.
• Make-to-stock environment.
• Dedicated production lines.
• Production lines that each produces two or more
families of products.
• Large setup times and setup costs for families, as
opposed to low setup times and costs for individual
items.
The two primary objectives of the production system
as described by the authors are to smooth peak demands through time so as not to exceed production capacity and to allocate production requirements among
the families to balance family holding and setup costs.
The planning is done with a six-month time horizon for
demand forecasting. The six-month period is divided
into two portions: the next four weeks and the remaining five months. Detailed plans are developed for the
near term, including regular and overtime production
allocation.
The model has two primary components: a family
planning model, which finds the optimal timing and sizing of family production runs, and a disaggregation
planning model, which takes the results of family planning and determines lot sizes for individual items within
families.
The authors also discuss several implementation issues specific to Welch’s environment. Product run
lengths must be tied to the existing eight-hour shift
structure. To do so; they recommend that production
run lengths be expressed as multiples of one-quarter
shift (two hours).
The model was implemented on a personal computer.
Computing times are very moderate. Solution techniques include a mixed integer mathematical program
and a linear programming formulation with relaxation
(that is, rounding of variables to integer values).
This case demonstrates that the concepts discussed in
this chapter can be useful in a real production planning
environment. Although the system described here is not
based on any of the specific models discussed in this chapter, this application shows that aggregation and disaggregation are useful concepts. Hierarchical aggregation
for production scheduling is a valuable planning tool.
The concept of disaggregating the aggregate plan along organizational lines in a fashThethat
upper
and lower with
bounds
Yj accountscheme
for possible
constraints
on the producion
is consistent
theon
aggregation
is an side
appealing
one. Whether
or not
tionmethods
level fordiscussed
item j. in this section provide a workable link between aggregate plans
the
numberitem
of feasibility
need
to be
addressed before the family run sizes Yj
andAdetailed
schedulesissues
remains
to be
seen.
areAnother
further disaggregated
intodisaggregation
lots for individual
items.has
The
objective
is tobyschedule
approach to the
problem
been
explored
Chung
the lots
for individual
items
within
a family
so that they programming
run out at the scheduled
setup
and
Krajewski
(1984).
They
develop
a mathematical
formulation
of
timeproblem.
for the family.
way, items
withinaggregate
the same family
caneach
be produced
within
the
Inputs In
to this
the program
include
plans for
product family.
the same
production
setup.setup status, total production level for the family, invenThis
includes
setup time,
concept
of disaggregating
the aggregate
plan
along
organizational
lines
toryThe
level,
workforce
level, overtime,
and regular
time
availability.
The goal
ofin
thea
fashion that
is consistent
theand
aggregation
is an appealing
one. Whether
or
analysis
is to
specify lotwith
sizes
timing ofscheme
production
runs for each
individual
not theconsistent
methods discussed
in this section
provide afor
workable
link between
aggregate
item,
with the aggregate
information
the product
family. Although
plans aand
detailed item
schedules
remains to
be between
seen.
such
formulation
provides
a potential
link
the aggregate plan and the
Another
approachschedule,
to the disaggregation
has beenprogram
exploredrequires
by Chung
and
master
production
the resultingproblem
mathematical
many
(1984).
They
a mathematical
formulation
the
­iKrajewski
nputs and can
result
in a develop
very large
mixed integer programming
problem that could
be veryof
timeproblem. Inputs
to the program include aggregate plans for each product family. This
consuming
to solve.
includes setup time, setup status, total production level for the family, inventory level,
workforce level, overtime, and regular time availability. The goal of the analysis is to
specify lot sizes and timing of production runs for each individual item, consistent with
the aggregate information for the product family. Although such a formulation
nah77856_ch03.indd 159
17/12/14 6:48 PM
160 Chapter Three Sales and Operations Planning
Problems for Section 3.6
29. What does “disaggregation of aggregate plans” mean?
30. Discuss the following quotation made by a production manager: “Aggregate
planning is useless to me because the results have nothing to do with my master
­production schedule.”
3.7 sales and operation Planning on a Global Scale
Globalization of manufacturing operations is commonplace. Many major corporations are now classified as multinationals; manufacturing and distribution ­activities
routinely cross international borders. With the globalization of both sources of
­
production and markets, firms must rethink production planning strategies. One
issue explored in this chapter was smoothing of production plans over time; costs
of ­increasing or ­decreasing workforce levels (and, hence, production levels) play
a major role in the optimization of any aggregate plan. When formulating global
production strategies, other smoothing issues arise. Exchange rates, costs of direct
labor, and tax structure are just some of the differences among countries that must be
factored into a global strategy.
Why the increased interest in global operations? In short, cost and competitiveness.
According to McGrath and Bequillard (1989):
The benefits of a properly executed international manufacturing strategy can be very substantial. A well developed strategy can have a direct impact on the financial performance
and ultimately be reflected in increased profitability. In the electronic industry, there are
examples of companies attributing 5% to 15% reduction in cost of goods sold, 10% to 20%
increase in sales, 50% to 150% improvement in asset utilization, and 30% to 100% increase
in inventory turnover to their internationalization of manufacturing.
Cohen et al. (1989) outline some of the issues that a firm must consider when planning
production levels on a worldwide basis. These include
• In order to achieve the kinds of economies of scale required to be competitive
today, multinational plants and vendors must be managed as a global system.
• Duties and tariffs are based on material flows. Their impact must be factored into
decisions regarding shipments of raw material, intermediate product, and finished
product across national boundaries.
• Exchange rates fluctuate randomly and affect production costs and pricing decisions in countries where the product is produced and sold.
• Corporate tax rates vary widely from one country to another.
• Global sourcing must take into account longer lead times, lower unit costs, and
­access to new technologies.
• Strategies for market penetration, local content rules, and quotas constrain product
flow across borders.
• Product designs may vary by national market.
• Centralized control of multinational enterprises creates difficulties for several reasons, and decentralized control requires coordination.
• Cultural, language, and skill differences can be significant.
nah77856_ch03.indd 160
17/12/14 6:48 PM
3.8 Historical Notes 161
Determining optimal globalized manufacturing strategies is clearly a daunting problem
for any multinational firm. One can formulate and solve mathematical models similar
to the linear programming formulations of the aggregate planning models ­presented in
this chapter, but the results of these models must always be balanced against judgment
and experience. Cohen et al. (1989) consider such a model. They assume multiple
products, plants, markets, raw materials, vendors, vendor supply contract ­alternatives,
time periods, and countries. Their formulation is a large-scale mixed integer, nonlinear
program.
One issue not treated in their model is that of exchange rate fluctuations and their
effect on both pricing and production planning. Pricing, in particular, is traditionally
done by adding a markup to unit costs in the home market. This completely ignores
the issue of exchange rate fluctuations and can lead to unreasonable prices in some
countries. For example, this issue has arisen at Caterpillar Tractor (Caterpillar ­Tractor
Company, 1985). In this case, dealers all over the world were billed in U.S. ­dollars
based on U.S. production costs. When the dollar was strong relative to other currencies, retail prices charged to overseas customers were not competitive in local markets.
Caterpillar found itself losing market share abroad as a result. In the early 1980s the
firm switched to a locally competitive pricing strategy to counteract this problem.
The notion that manufacturing capacity can be used as a hedge against exchange
rate fluctuations has been explored by several researchers. Kogut and Kulatilaka
(1994), for example, develop a mathematical model for determining when it is optimal to switch production from one location to another. Since the cost of switching is
­assumed to be positive, there must be a sufficiently large difference in exchange rates
before switching is recommended. As an example, they consider a situat­ion where a
firm can produce its product in either the United States or Germany. If production is
currently being done in one location, the model provides a means of determining if it
is economical to switch locations based on the relative strengths of the euro and dollar. While such models are in the early stages of development, they provide a means
of ­rationalizing international production planning strategies. Similar issues have been
­explored by Huchzermeier and Cohen (1996) as well.
3.8 Historical Notes
The aggregate planning problem was conceived in an important series of papers that
­appeared in the mid-1950s. The first, Holt, Modigliani, and Simon (1955), discussed
the structure of the problem and introduced the quadratic cost approach, and the later
study of Holt, Modigliani, and Muth (1956) concentrated on the computational aspects
of the model. A complete description of the method and its application to production
planning for a paint company is presented in Holt, Modigliani, Muth, and Simon (1960).
It should be recognized that the text by Holt et al. (1960) represents a landmark
work in the application of quantitative methods to production planning problems. The
authors developed a solution method that results in a set of formulas that are easy to
implement and they actually undertook the implementation of the method. The work
details the application of the approach to a large manufacturer of household paints in
the Pittsburgh area. The analysis was implemented in the company but a subsequent
visit to the firm indicated that serious problems arose when the linear decision rule was
followed, primarily because of the firm’s policy of not firing workers when the model
indicated that they should be fired.
nah77856_ch03.indd 161
17/12/14 6:48 PM
162 Chapter Three Sales and Operations Planning
That production planning problems could be formulated as linear programs appears to
have been known in the early 1950s. Bowman (1956) discussed the use of a transportation model for production planning. The particular linear programming formulation of
the ­aggregate planning problem discussed in Section 3.5 is essentially the same as the one
­developed by Hansmann and Hess (1960). Other linear programming formulations of the
production planning problem generally involve multiple products or more complex cost
structures (see, for example, Newson, 1975a and 1975b).
More recent work on the aggregate planning problem has focused on aggregation
and disaggregation issues (Axsater, 1981; Bitran and Hax, 1981; and Zoller, 1971), the
incorporation of learning curves into linear decision rules (Ebert, 1976), extensions to
allow for multiple products (Bergstrom and Smith, 1970), and inclusion of marketing
and/or financial variables (Damon and Schramm, 1972, and Leitch, 1974).
Taubert (1968) considers a technique he refers to as the search decision rule. The
method requires developing a computer simulation model of the system and searching the response surface using standard search techniques to obtain a (not necessarily
optimal) solution. Taubert’s approach, which was described in detail in Buffa and Taubert (1972), gives results that are comparable to those of Holt, Modigliani, Muth, and
Simon (1960) for the case of the paint company.
Kamien and Li (1990) developed a mathematical model to examine the effects of
subcontracting on aggregate planning decisions. The authors show that under certain
circumstances it is preferred to producing in-house, and provides an additional means
of smoothing production and workforce levels.
3.9 Summary
nah77856_ch03.indd 162
Modern firms take a much less siloed approach to planning than firms of the past. They
have come to realize that large benefits can accrue from a collaborative approach among the
different divisions including operations, sales and marketing, and finance. A well ­designed
S&OP process can manage the inherent tensions between these divisions and make tradeoffs from a strategic perspective. The key output from an operational perspective of this
process is a fixed sales plan that the operations division can plan to.
A common component of the S&OP process will be reviewing divisional KPIs. Because
KPIs have such a large impact on employee incentives they need to be carefully chosen.
The primary challenge in KPI selection is to ensure that the KPI is aligned with the strategic imperatives of the firm while still remaining actionable as far as the person being
measured by it is concerned. It is also important to try to mitigate opportunities for gaming
of KPIs by employees.
The key output from the S&OP process is a fixed forecast and therefore planning for
routine uncertainty in these numbers is typically left to the operations division. While risk
pooling (see Chapter 6) can be used to mitigate some of the uncertainty, there will always
be natural variation that must be buffered. Such buffering can take the form of inventory,
spare capacity, or time in the form of customer lead times. This sort of buffering typically
takes place outside the S&OP process and will be discussed in later chapters in this text.
At a higher level than routine variation are more major risks that are discussed within
the S&OP process. Such known unknowns should have contingency plans and/or mitigation strategies associated with them. They may be demand-side risks, such as new product
introductions by competitors or demand shocks caused by extreme weather, or they may
be supply-side risks such as supplier failure or major quality issues. An effective S&OP
can anticipate such risks and set strategies for dealing with them that are in line with the
company strategy. This may include the explicit relaxing of certain KPIs that become less
relevant when working under exceptional circumstances.
17/12/14 6:48 PM
Additional Problems on Aggregate Planning 163
Determining optimal production levels for all products produced by a large firm can be an
enormous undertaking. Aggregate planning addresses this problem by assuming that individual
items can be grouped together. However, finding an effective aggregating scheme can be difficult and often revenue dollars are used for simplicity. One particular aggregating scheme that
has been suggested is items, families, and types. Items (or stock keeping units, SKUs), represent the finest level of detail, are identified by separate part numbers, bar codes, and/or radio
frequency ID (RFID) tags when appropriate. Families are groups of items that share a common manufacturing setup, and types are natural groups of families. This particular aggregation
scheme is fairly general but there is no guarantee that it will work in every application.
As mentioned above, once the strategies for dealing with uncertainty have been determined, most firms assume determinist demand in coming up with an aggregate production
plan. Indeed, fixed demand forecasts over a specified planning horizon are required input.
This assumption is not made for the sake of realism, but to allow the analysis to focus on
the changes in the demand that are systematic rather than random. The goal of the analysis
is to determine for each period the number of workers that should be employed, the production that should occur, and the inventory that should be carried over for each period.
The objective of an aggregate production plan is to minimize costs of production, payroll,
holding, and changing the size of the workforce or capacity. The costs of making changes
are generally referred to as smoothing costs. The aggregate planning models discussed in this
chapter assume that all the costs are linear functions. This assumption is probably a reasonable approximation for most real systems within a given reasonable range of values. It is
unlikely that the primary problem with applying a linear programming formulation to a real
situation will be that the shape of the cost function is incorrect; it is more likely that the primary difficulty will be in correctly estimating the costs and other required input information.
Aggregate production plans will be of little use to the firm if they cannot be coordinated with detailed item schedules (i.e., the master production schedule). The problem
of ­disaggregating aggregate plans is a difficult one, but one that must be addressed if
the aggregate plans are to have value to the firm. There have been some mathematical
­programming formulations of the disaggregating problem suggested in the literature but
these disaggregation schemes have yet to be proven in practice.
Additional
Problems on
Aggregate
Planning
31. An aggregate planning model is being considered for the following applications.
­Suggest an aggregate measure of production and discuss the difficulties of applying
aggregate planning in each case.
a. Planning for the size of the faculty in a university.
b. Determining workforce requirements for a travel agency.
c. Planning workforce and production levels in a fish-processing plant that ­pro­duces
canned sardines, anchovies, kippers, and smoked oysters.
32. A local firm manufactures children’s toys. The projected demand over the next four
months for one particular model of toy robot is
Month
Workdays
July
August
September
October
nah77856_ch03.indd 163
23
16
20
22
Forecasted Demand
(in aggregate units)
3,825
7,245
2,770
4,440
17/12/14 6:48 PM
164 Chapter Three Sales and Operations Planning
Assume that a normal workday is eight hours. Hiring costs are $350 per worker
and firing costs (including severance pay) are $850 per worker. Holding costs
are $4.00 per aggregate unit held per month. Assume that it requires an ­average
of 1 hour and 40 minutes for one worker to assemble one toy. Shortages are not
­permitted. ­Assume that the ending inventory for June was 600 of these toys and the
manager wishes to have at least 800 units on hand at the end of O
­ ctober. Assume that
the ­current workforce level is 35 workers. Find the optimal plan by formulating as a
linear program.
33. The Paris Paint Company is in the process of planning labor force requirements and
­production levels for the next four quarters. The marketing department has provided
­production with the following forecasts of demand for Paris Paint over the next year:
Quarter
1
2
3
4
Demand Forecast
(in thousands of gallons)
380
630
220
160
Assume that there are currently 280 employees with the company. ­Employees are
hired for at least one full quarter. Hiring costs amount to $1,200 per ­employee and firing costs are $2,500 per employee. Inventory costs are $1 per gallon per quarter. It is
­estimated that one worker produces 1,000 ­gallons of paint each ­quarter.
Assume that Paris currently has 80,000 gallons of paint in inventory and would like
to end the year with an inventory of at least 20,000 gallons.
a. Determine the minimum constant workforce plan for Paris Paint and the cost of the
plan. Assume that stock-outs are not allowed.
b. If Paris were able to back-order excess demand at a cost of $2 per gallon per quarter, determine a minimum constant workforce plan that holds less i­n­ventory than
the plan you found in part (a), but incurs stock-outs in quarter 2. Determine the cost
of the new plan.
c. Formulate this as a linear program. Assume that stock-outs are not allowed.
d. Solve the linear program. Round the variables and determine the cost of the resulting plan.
34. Consider the problem of Paris Paint presented in Problem 33. Suppose that the plant
has the capacity to employ a maximum of 370 workers. Suppose that ­regular-time
­employee costs are $12.50 per hour. Assume seven-hour days, five-day weeks, and
four-week months. Overtime is paid on a time-and-a-half basis. S
­ ubcontracting is
available at a cost of $7 per gallon of paint produced. Overtime is limited to three
hours per employee per day, and no more than 100,000 gallons can be subcontracted
in any quarter.
a. Formulate as a linear program.
b. Solve the linear program. Round the variables and determine the cost of the
­resulting plan.
nah77856_ch03.indd 164
17/12/14 6:48 PM
165
Additional Problems on Aggregate Planning 165
35. The Mr. Meadows Cookie Company can obtain accurate forecasts for 12 months based
on firm orders. These forecasts and the number of workdays per month are as follows:
Month
Demand Forecast
(in thousands of cookies)
Workdays
1
850
2
1,260
3
510
4
980
5
770
6
850
7
1,050
8
1,550
9
1,350
10
1,000
11 970
12 680
26
24
20
18
22
23
14
21
23
24
21
13
During a 46-day period when there were 120 workers, the firm produced
1,700,000 cookies. Assume that there are 100 workers employed at the beginning of
month 1 and zero starting inventory.
a. Find the minimum constant workforce needed to meet monthly demand.
b. Assume cI 5 $0.10 per cookie per month, cH 5 $100, and cF 5 $200. Add columns
that give the cumulative on-hand inventory and inventory cost. What is the total
cost of the constant workforce plan?
c. Solve for the optimal plan using linear programming. Compare your solution to b.
36. The Yeasty Brewing Company produces a popular local beer known as Iron Stomach.
Beer sales are somewhat seasonal, and Yeasty is planning its p­ roduction and ­workforce
levels on March 31 for the next six months. The ­demand forecasts are as follows:
Month
April
May
June
July
August
September
Production
Days
Forecasted Demand
(in hundreds of cases)
11 85
22 93
20
122
23
176
16
140
20 63
As of March 31, Yeasty had 86 workers on the payroll. Over a period of 26 working days when there were 100 workers on the payroll, Yeasty produced 12,000 cases
of beer. The cost to hire each worker is $125 and the cost of l­aying off each worker
is $300. Holding costs amount to 75 cents per case per month.
As of March 31, Yeasty expects to have 4,500 cases of beer in stock, and it wants
to maintain a minimum buffer inventory of 1,000 cases each month. It plans to start
­October with 3,000 cases on hand.
nah77856_ch03.indd 165
17/12/14 6:48 PM
166 Chapter Three Sales and Operations Planning
a. Based on this information, find the minimum constant workforce plan for Yeasty
over the six months, and determine hiring, firing, and holding costs ­associated with
that plan.
b. Suppose that it takes one month to train a new worker. How will that ­affect your
­solution?
c. Suppose that the maximum number of workers that the company can e­ xpect to
be able to hire in one month is 10. How will that affect your ­solution to part (a)?
d. F
ormulate the problem levels as a linear program. [You may ignore the conditions
in parts (b) and (c).]
e. Solve the resulting linear program. Round the appropriate variables and ­determine
the cost of your solution.
f. Suppose Yeasty does not wish to fire any workers. What is the optimal plan subject
to this constraint?
37. A local canning company sells canned vegetables to a supermarket chain in the ­Minneapolis
area. A typical case of canned vegetables requires an average of 0.2 day of labor to produce. The aggregate inventory on hand at the end of June is 800 cases. The demand for
the vegetables can be accurately predicted for about 18 months based on orders received
by the firm. The predicted demands for the next 18 months are as follows:
Month
July
August
September
October
November
December
January
February
March
Forecasted Forecasted
Demand Demand
(hundreds (hundreds
of cases)
Workdays
Month
of cases)
23
28
42
26
29
58
19
17
25
21
14
20
23
18
10
20
14
20
April
May
June
July
August
September
October
November
December
29
33
31
20
16
33
35
28
50
Workdays
20
22
21
18
14
20
23
18
10
The firm currently has 25 workers. The cost of hiring and training a new worker
is $1,000, and the cost to lay off one worker is $1,500. The firm estimates a cost of
$2.80 to store a case of vegetables for a month. They would like to have 1,500 cases in
­inventory at the end of the 18-month planning h­ orizon.
a. Develop a spreadsheet to find the minimum constant workforce aggregate plan and
determine the total cost of that plan.
b. Develop a spreadsheet to find a plan that hires and fires workers monthly in order
to minimize inventory costs. Determine the total cost of that plan as well.
nah77856_ch03.indd 166
17/12/14 6:48 PM
Appendix 3–A 167
Appendix 3–A
Glossary of Notation for Chapter 3
a 5 Smoothing constant for production and demand used in Bowman’s model.
b 5 Smoothing constant for inventory used in Bowman’s model.
cF 5 Cost of firing one worker.
cH 5 Cost of hiring one worker.
cI 5 Cost of holding one unit of stock for one period.
cO 5 Cost of producing one unit on overtime.
cP 5 Penalty cost for back orders.
cR 5 Cost of producing one unit on regular time.
cS 5 Cost to subcontract one unit of production.
cU 5 Idle cost per unit of production.
Dt 5 Forecast of demand in period t.
Ft 5 Number of workers fired in period t.
Ht 5 Number of workers hired in period t.
It 5 Inventory level in period t.
K 5 Number of aggregate units produced by one worker in one day.
lj 5 Annual demand for family j (refer to Section 3.9).
nt 5 Number of production days in period t.
Ot 5 Overtime production in units.
Pt 5 Production level in period t.
St 5 Number of units subcontracted from outside.
T 5 Number of periods in the planning horizon.
Ut 5 Worker idle time in units (“undertime”).
Wt 5 Workforce level in period t.
Bibliography
ABB Group. “Key Performance Indicators: Identifying and
using key metrics for performance.” PowerPoint accessed
from http://www.abb.com/Search.aspx?q=pr&abbcontext=pr
oducts&num=10&filetype=mspowerpoint&filter=0&start=10
Allen, S. J., and E. W. Schuster. “Practical Production
­Scheduling with Capacity Constraints and Dynamic D
­ emand:
Family Planning and Disaggregation.” Production and
­Inventory Management Journal 35 (1994), pp. 15–20.
Axsater, S. “Aggregation of Product Data for Hierarchical
­Production Planning.” Operations Research 29 (1981),
pp. 744–56.
nah77856_ch03.indd 167
Bergstron, G. L., and B. E. Smith. “Multi-Item Production
Planning—An Extension of the HMMS Rules.” Management
Science 16 (1970), pp. 100–103.
Bitran, G. R., and A. Hax. “Disaggregation and Resource
Allocation Using Convex Knapsack Problems with
Bounded Variables.” Management Science 27 (1981),
pp. 431–41.
Buffa, E. S., and W. H. Taubert. Production-Inventory
­Systems: Planning and Control. Rev. ed. New York:
McGraw-Hill/Irwin, 1972.
17/12/14 6:48 PM
168 Chapter Three Sales and Operations Planning
Chopra, S. and P. Meindl. Supply Chain Management: Strategy, Planning & Operation. 3rd ed. Upper Saddle River, NJ:
Prentice Hall, 2006.
Chung, C., and L. J. Krajewski. “Planning Horizons for
­Master Production Scheduling.” Journal of Operations
­Management (1984), pp. 389–406.
De Meyer, A., C. H. Loch, and M. T. Pich. “Managing
­Project Uncertainty: From Variation to Chaos.” MIT Sloan
Management Review (2002).
Damon, W. W., and R. Schramm. “A Simultaneous
­Decision Model for Production, Marketing, and Finance.”
­Management Science 19 (1972), pp. 16 –72.
Ebert, R. J. “Aggregate Planning with Learning Curve
Productivity.” Management Science 23 (1976), pp. 171–82.
Erenguc, S., and S. Tufekci. “A Transportation Type
­Aggregate Production Model with Bounds on Inventory and
Backordering.” European Journal of Operations Research 35
(1988), pp. 414–25.
EveryAngle Software.Accessed from http://www.everyangle.
com/downloads/customer-cases/en/customercase_heineken.pdf
Goldratt, E. M. and J. Cox. The Goal: A Process of Ongoing
Improvement. 2nd Revised Edition. Great Barrington, MA:
North River Press, 1992.
Grimson, J. A., and D. F. Pyke. “Sales and operations planning: an exploratory study and framework.” International
Journal of Logistics Management 18 (2007), pp.322–346.
Hansmann, F., and S. W. Hess. “A Linear Programming
­Approach to Production and Employment Scheduling.”
­Management Technology 1 (1960), pp. 46–51.
Hax, A. C., and D. Candea. Production and Inventory
­Management. Englewood Cliffs, NJ: Prentice Hall, 1984.
Hax, A. C., and H. C. Meal., “Hierarchical Integration of
Production Planning and Scheduling.” In TIMS Studies in
Management Science. Volume 1, Logistics, ed. M. Geisler.
New York: Elsevier, 1975.
Hiller, F. S., and G. J. Lieberman. Introduction to Operations
Research. 5th ed. San Francisco: Holden Day, 1990.
Holt, C. C., F. Modigliani, and J. F. Muth. “Derivation of
a Linear Decision Rule for Production and Employment.”
Management Science 2 (1956), pp. 159–77.
Holt, C. C., F. Modigliani, J. F. Muth, and H. A. Simon.
Planning Production, Inventories, and Workforce. Englewood Cliffs, NJ: Prentice Hall, 1960.
Holt, C. C., F. Modigliani; and H. A. Simon. “A Linear
­Decision Rule for Employment and Production Scheduling.”
Management Science 2 (1955), pp. 1–30.
nah77856_ch03.indd 168
Hopp, W. J., and M. L. Spearman. Factory Physics. (1996).
Boston, MA: McGraw Hill.
IBM. “IBM Global Chief Supply Chain Officer Study.”
­Accessed from http://www-935.ibm.com/services/us/gbs/bus/
html/gbs-csco-study.html
Jordan, W. C., and S. C. Graves. “Principles on the benefits
of manufacturing process flexibility.” Management Science
41 (1995), pp. 577–594.
Kaplan, R. S., and D. P. Norton. “The Balanced Scorecard:
Measures that Drive Performance.” Harvard Business
­Review (1992), pp. 71–80.
Kamien, M. I., and L. Li. “Subcontracting, Coordination, and
Flexibility, and Production Smoothing Aggregate Planning.”
Management Science 36 (1990), pp. 1352–63.
Leitch, R. A. “Marketing Strategy and Optimal Production Schedule.” Management Science 20 (1974),
pp. 903–11.
Newson, E. F. P. “Multi-Item Lost Size Scheduling by
­Heuristic, Part 1: With Fixed Resources.” Management
­Science 21 (1975a), pp. 1186–93.
Newson, E. F. P. “Multi-Item Lost Size Scheduling by
­Heuristic, Part 2: With Fixed Resources.” Management
­Science 21 (1975b), pp. 1194–1205.
Rooijen, H. “Connecting our supply chain to our customers. Accessed from http://www.heinekeninternational.com/
content/live/files%202011/investors/6.%20Henk%20van%20
Rooijen.pdf
Santos, C., et al. “HP Enterprise Services Uses Optimization for Resource Planning.” Interfaces 43 (2013),
pp. 152–169.
Sheldon, D. H. World Class Sales & Operations Planning.
Ft. Lauderdale, FL: J Ross Publishing and APICS, 2006.
Smits, J. and M. English. “The Journey to Worldclass S&OP
at Heineken.” Accessed from http://supply-chain.org/
node/17283
Taleb, N. N. The Black Swan: The Impact of the Highly
Improbable. New York: Random House, 2007.
Taubert, W. H. “A Search Decision Rule for the Aggregate
Scheduling Problem.” Management Science 14 (1968),
pp. B343–53.
Vollmann, T. E., W. L. Berry, and D. C. Whybark. Manufacturing, Planning, and Control Systems. 3rd ed. New York:
McGraw-Hill/Irwin, 1992.
Zoller, K. “Optimal Disaggregation of Aggregate Production
Plans.” Management Science 17 (1971), pp. B53–49.
17/12/14 6:48 PM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 169
Supplement One
Linear Programming
S1.1
INTRODUCTION
Linear programming is a mathematical technique for solving a broad class of optimization problems. These problems require maximizing or minimizing a linear function of n real variables subject to m constraints. One can formulate and solve a large
number of real problems with linear programming. A partial list includes
1. Scheduling of personnel.
2. Several varieties of blending problems including cattle feed, sausages, ice cream,
and steel making.
3. Inventory control and production planning.
4. Distribution and logistics problems.
5. Assignment problems.
Problems with thousands of variables and thousands of constraints are easily solvable
on computers today. Linear programming was developed to solve logistics problems
during World War II. George Dantzig, a mathematician employed by the RAND Corporation at the time, developed a solution procedure he labeled the Simplex Method. That
the method turned out to be so efficient for solving large problems quickly was a surprise even to its developer. That fact, coupled with the simultaneous development of the
electronic computer, established linear programming as an important tool in logistics
management. The success of linear programming in industry spawned the development
of the disciplines of operations research and management science. The Simplex Method
has withstood the test of time. Only in recent years has another method been developed
that potentially could be more efficient than the Simplex Method for solving very large,
specially structured, linear programs. This method, known as Karmarkar’s Algorithm, is
named for the Bell Labs mathematician who conceived it.
In Section S1.2 we consider a typical manufacturing problem that we formulate and
solve using linear programming. Later we explore how one solves small problems (that
is, having exactly two decision variables) graphically, and how one solves large
problems using a computer.
S1.2
A PROTOTYPE LINEAR PROGRAMMING PROBLEM
Example S1.1
Sidneyville manufactures household and commercial furnishings. The Office Division produces
two desks, rolltop and regular. Sidneyville constructs the desks in its plant outside Medford,
Oregon, from a selection of woods. The woods are cut to a uniform thickness of 1 inch. For this
reason, one measures the wood in units of square feet. One rolltop desk requires 10 square feet
of pine, 4 square feet of cedar, and 15 square feet of maple. One regular desk requires
169
169
nah77856_ch03Supple.indd 169
17/12/14 11:58 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 170
170 Supplement
One Linear
170 Supplement One
Linear Programming
20 square feet of pine, 16 square feet of cedar, and 10 square feet of maple. The desks yield
respectively $115 and $90 profit per sale. At the current time the firm has available 200 square
feet of pine, 128 square feet of cedar, and 220 square feet of maple. The firm has backlogged
orders for both desks and would like to produce the number of rolltop and regular desks that
would maximize profit. How many of each should it produce?
Solution
The first step in formulating a problem as a linear program is to identify the decision variables. In
this case there are two decisions required: the number of rolltop desks to produce and the number of regular desks to produce. We must assign symbol names to each of these decision variables.
Let
x1 � Number of rolltop desks to be produced,
x2 � Number of regular desks to be produced.
Now that we have identified the decision variables, the next step is to identify the objective
function and the constraints. The objective function is the quantity we wish to maximize or
minimize. The objective is to maximize the profits, so the objective function equals the total
profit when producing x1 rolltop desks and x2 regular desks. Each rolltop desk contributes $115
to profit, so the total contribution to profit from all rolltop desks is 115x1. Similarly, the contribution to profit from all regular desks is 90x2. Hence, the total profit is 115x1 � 90x2. This
is known as the objective function.
The next step is to identify the constraints. The number of desks Sidneyville can produce is
limited by the amount of wood available. The three types of wood constitute the critical
resources. To obtain the constraints, we need to find expressions for the amount of each type
of wood consumed by construction of x1 rolltop desks and x2 regular desks. Those expressions
are then bounded by the amount of each type of wood available.
The number of square feet of pine used to make x1 rolltop desks is 10x1. The number
of square feet of pine used to make x2 regular desks is 20x2. It follows that the total amount of
pine consumed in square feet is 10x1 � 20x2. This quantity cannot exceed the number
of square feet of pine available, which is 200. Hence we obtain the first constraint:
10x1 � 20x2 � 200.
The other two constraints are similar. The second constraint is to ensure that the firm does
not exceed the available supply of cedar. Each rolltop desk requires 4 square feet of cedar, so
x1 rolltop desks require 4x1 square feet of cedar. Each regular desk requires 16 square feet of
cedar, so x2 regular desks require 16x2 square feet of cedar. It follows that the constraint on the
supply of cedar is
4x1 � 16x2 � 128.
In the same way, the final constraint ensuring that we do not exceed the supply of maple is
15x1 � 10x2 � 220.
Because we cannot produce a negative number of desks, we also include nonnegativity
constraints:
x1 � 0,
x2 � 0.
We have now constructed the complete linear programming formulation of the Sidneyville
problem. The goal is to find values of x1 and x2 to maximize 115x1 � 90x2, subject to the
constraints.
10x1 � 20x2 � 200,
4x1 � 16x2 � 128,
15x1 � 10x2 � 220,
x1, x2 � 0.
nah77856_ch03Supple.indd 170
17/12/14 11:58 AM
nah77856_ch03Supple.qxd
1/29/08
5:15 PM
Page 171
S1.3 Statement of
of the
the General
General Problem 171
Problem 171
S1.3 Statement
TABLE S1–1 Partial Computer Output for Sidneyville Problem
LP OPTIMUM FOUND AT STEP 2
OBJECTIVE FUNCTION VALUE
1) 1740.000000
VARIABLE
X1
X2
VALUE
12.000000
4.000000
REDUCED COST
.000000
.000000
We next consider how such a problem is solved. We will briefly outline the theory behind
the solution technique known as the Simplex Method. However, as linear programming problems are almost never solved by hand any longer, you will not need to understand the mechanics of the Simplex Method in order to use linear programming. You will need to know how
to formulate problems as linear programs, enter the formulations into the computer, recognize
special problems, and analyze the computer’s output.
A typical computer output is given in Table S1–1.
This output tells us that at the optimal solution, the value of x1 is 12 and the value of x2 is 4.
That is, Sidneyville should produce exactly 12 rolltop desks and 4 regular desks. The value of the
objective function at the optimal solution is $1,740.00. That this is the (unique) optimal solution means the following: every other pair of values of x1 and x2 will result in either a lower
profit, not meeting the constraints, or both.
Sidneyville’s manager of production planning is very skeptical about mathematics. When
presented with this solution, his response was, “There’s only one problem with this production
plan. We have a specialist make the rolltop portion of the rolltop desk. She can only do four
desks a day and we want to be ready to ship out in two days. There is no way we can produce
twelve of those desks in two days. I knew this math stuff was a lot of hooey!”
The manager was wrong. The trouble is not that the formulation is incorrect, but that it does
not include all relevant constraints, as labor hours turned out to be a critical resource. The
lesson here is that for the final solution to be meaningful, the model must include all relevant
constraints.
S1.3
STATEMENT OF THE GENERAL PROBLEM
The Sidneyville problem is an example of a linear program in which there are two
decision variables and three constraints. Linear programming problems may have any
number of decision variables and any number of constraints. Suppose that there are
n decision variables, labeled x1, x2, . . . , xn, subject to m resource constraints. Then we
may write the problem of maximizing the objective subject to the constraints as
Maximize c1x1 � c2x2 � . . . � cnxn ,
subject to a11x1 � a12x2 � . . . � a1n xn � b1,
a21x1 � a 22x2 � . . . � a2n xn � b2,
.
.
.
a m1x1 � am2x2 � . . . � amn xn � bm,
x1, x2, . . . , xn � 0.
Interpret c1, c2, . . . , cn as the profit coefficients per unit of output of the activities
x1, x2, . . . , xn; aij as the amount of resource i consumed by one unit of activity j; and
nah77856_ch03Supple.indd 171
17/12/14 11:58 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 172
nah77856
172
Supplement One
Linear Programming
172 Supplement
One Linear
Programming
b1 as the amount of resource i available, for i � 1, . . . , m and j � 1, . . . , n. We require
that the constants b1, . . . , bm be nonnegative. This particular formulation includes
problems in which we want to maximize profit subject to constraints on the available
resources. However, linear programming can be used to solve a much larger variety of
problems. Other possible formulations will be discussed later.
Definitions of Commonly Used Terms
1. Objective function. This is the quantity we wish to maximize or minimize. In the
given formulation, the objective function is the term c1x1 � c2x2 � . . . � cnxn. In business applications, one typically minimizes cost or maximizes profit. We use the abbreviations “min” for a minimization problem and “max” for a maximization problem.
2. Constraints. Each constraint is a linear inequality or equation, that is, a linear
combination of the problem variables followed by a relational operator (� or � or �)
followed by a nonnegative constant. Although the given formulation shows all � constraints, � and � type constraints are also common. For example, suppose there is a
contractual agreement that requires a minimum number of labor hours daily. This
would result in a � constraint.
3. Right-hand side. The right-hand side is the constant following the relational operator in a constraint. In the given constraint, the constants b1, b2, . . . , bm are the righthand sides. These constants are required to be nonnegative numbers. However, we do
not require that the constants aij be nonnegative. This means that any constraint can be
written with a nonnegative right-hand side by multiplying through by the constant �1
whenever the right-hand side is negative. Consider the following simple example.
Suppose that when formulating a problem as a linear program we obtain the constraint
4x1 � 2x2 � �5.
Because the right-hand side is negative, this is not a legal constraint. However, if we
multiply through by �1, this constraint becomes
�4x1 � 2x2 � 5,
which is acceptable.
4. Feasible region. The feasible region is the set of values of the decision variables,
x1, x2, . . . , xn, that satisfy the constraints. Because each of the constraints is generated
by a linear equation or linear inequality, the feasible region has a particular structure.
The technical term for this structure is convex polytope. In two dimensions, a convex
polytope is a convex set with boundaries that are straight lines. In three dimensions, the
boundaries are formed by planes. A convex set is characterized as follows: pick any
two points in the set and connect them with a straight line; the line lies entirely within
the set.
5. Extreme points. Because of the structure of the feasible region, there will be a finite number of feasible points, with the property that they cannot be expressed as a linear combination of any other set of feasible points. These points are known as extreme
points or corner points, and they play an important role in linear programming. The concept of extreme points will become clearer when we discuss graphical solutions.
6. Feasible solution. A feasible solution is one particular set of values of the decision variables that satisfies the constraints. A feasible solution is also one point in the
feasible region. It may be an extreme point or an interior point.
7. Optimal solution. The optimal solution is the feasible solution that maximizes or
minimizes the objective function. In some cases the optimal solution may not be
unique. When this is the case, there will be an infinite number of optimal solutions.
nah77856_ch03Supple.indd 172
17/12/14 11:58 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 173
S1.3 Statement of
of the
the General
General Problem 173
Problem 173
S1.3 Statement
Features of Linear Programs
Linear programming is a very powerful tool. Many real problems have been successfully formulated and solved using this technique. However, to use the method correctly,
one must be aware of its limitations. Two important features of linear programs are
linearity and continuity. Many problems that may appear to be solvable by linear
programming fail one or both of these two crucial tests.
Linearity
Optimization problems can be formulated as linear programs only when (a) the objective can be expressed as a linear function of the decision variables and (b) all constraints can be expressed as linear functions of the decision variables.
Linearity implies that quantities change in fixed proportions. For example, if it costs
$10 to produce one unit, then it costs $20 to produce two units, and $100 to produce ten
units. If one ounce of orange juice supplies 30 mg of vitamin C, then three ounces must
supply 90 mg. Linearity must hold in the objective function and the constraints. In the objective function, this means that the profit or cost per unit must be the same independent
of the number of units. In the constraints, linearity means that the amount of each resource
consumed is the same per unit whether one produces a single unit or many units.
However, one often observes nonlinear relationships in the real world. Economies
of scale in production mean that the marginal cost of producing a unit decreases as the
number of units produced increases. When this occurs, the cost of production is a
nonlinear function of the number of units produced. An example of scale economies is
a fixed setup cost for production. The formula for the EOQ discussed in Chapter 4 says
that the lot size increases as the square root of the demand rate. Hence, EOQ is a
nonlinear function of the demand. When either the objective function or a constraint is
a nonlinear function of the decision variables, the problem is a nonlinear programming
problem and cannot be solved by linear programming.1
Continuity
This means that the decision variables should be continuous (that is, able to assume any
nonnegative value) as opposed to discrete or integer valued. This can be a serious restriction. The solution to many problems makes sense only if the decision variables are
integer valued. In particular, Example S1.1 is, strictly speaking, not a linear programming problem because the number of desks produced must be integer valued. (We were
lucky that the optimal solution turned out to be integer valued in this case.) One might
think that the optimal integer solution is equal to the continuous solution rounded off
to the nearest integer. Unfortunately, this is not always the case. First, rounding may
lead to infeasibility; that is, the rounded-off solution may lie outside the feasible
region. Second, even if the rounded solution is feasible, it may not be optimal. It can
happen that the optimal integer solution is in an entirely different portion of the feasible region than the rounded-off linear programming solution!
To give the reader some idea of the difficulties that can arise when the solution must
be integer valued, consider Example S1.1. Suppose that the profit from selling rolltop
desks was $150 rather than $115. Then the objective would be to maximize 150x1 90x2 subject to the same set of constraints. The optimal linear programming solution is
x1 14.666666 . . . .
x2 0.
1
In some circumstances, convex programming problems can be solved by linear programming by
approximating the objective function with a piecewise-linear function. An example of this approach
appears at the end of Section 3.5.
nah77856_ch03Supple.indd 173
17/12/14 11:58 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 174
nah77856
174
Supplement One
Linear Programming
174 Supplement
One Linear
Programming
Rounding the solution to the nearest integer gives x1 15 and x2 0, which is
infeasible. Substituting these values into the final constraint results in a requirement of
225 square feet of maple. However, only 220 square feet are available. Rounding x1
down to 14 results in a feasible but suboptimal solution. At x1 14 and x2 0, there
are 10 feet of maple still available, which is enough to make one regular desk. The
optimal integer solution in this case is x1 14 and x2 1.
When the decision variables must be integer valued, we say that the problem is an integer linear programming problem. Finding optimal integer solutions to linear programs
can be very time-consuming, even for modest-sized problems. However, Excel does
offer an option for defining some or all of the problem variables as integer valued. Excel
does a fine job solving small integer linear programming problems. For larger problems,
one would use a computer program designed to solve integer programming problems. In
some cases, and especially when the values of the variables are relatively large, careful
rounding of the linear programming solution should give acceptable results.
S1.4
SOLVING LINEAR PROGRAMMING PROBLEMS GRAPHICALLY
Graphing Linear Inequalities
In this section we will show how to solve two-variable linear programming problems
graphically. Although most real problems have more than two variables, understanding
the procedure for solving two-variable linear programs will improve your grasp of the
concepts underlying the Simplex Method.
The first step is to graph the linear inequalities represented by the constraints. A linear inequality corresponds to all points in the plane on one side of a straight line. There
are thus two steps to graphing linear inequalities:
1. Draw the straight line representing the boundary of the region corresponding to the
constraint expressed as an equation.
2. Determine which side of the line corresponds to the inequality.
To illustrate the method, consider the first constraint in Example S1.1.
10x1 20x2 200.
The boundary of the region represented by this inequality is the straight line
10x1 20x2 200.
The easiest way to graph a straight line is to determine the two intercepts. These are
found by setting x2 to zero and solving for x1, and then setting x1 to zero and solving
for x2. First setting x2 to zero gives
10x1 200
or
x1 20.
Similarly, setting x1 to zero and solving for x2 gives
20x2 200
or
x2 10.
Hence, the line 10x1 20x2 200 must pass through the points (20, 0) (the x1 intercept) and (0, 10) (the x2 intercept). A graph of this line is shown in Figure S1–1. Now
that we have graphed the line defining the boundary of the half space, we must determine which side of the line corresponds to the inequality. To do so, we pick any point
not on the line, substitute the values of x1 and x2, and see if the inequality is satisfied or
nah77856_ch03Supple.indd 174
17/12/14 11:58 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 175
S1.4 Solving Linear
Linear Programming
Programming Problems
Problems Graphically 175
Graphically 175
S1.4 Solving
FIGURE S1–1
Graphing a constraint
boundary
10
x2
10
x1
20
10x1 + 20x2 = 200
FIGURE S1–2
Half space
representing
the inequality
10x1 20x2 200
10
x2
x1
10
20
10x1 + 20x2 = 200
not. If the inequality is satisfied, then that point belongs in the half space; if it is not,
then that point does not belong in the half space. If the boundary line does not go
through the origin [that is, the point (0, 0)], then the most straightforward approach is
to use the origin as the point to be tested.
Substituting (0, 0) into the inequality, we obtain
(10)(0) (20)(0) 0 200.
Because 0 is less than 200, the test is satisfied and the origin lies within the region
represented by the inequality. This means that the graph of the inequality includes all
the points below the boundary line in Figure S1–1, as shown in Figure S1–2.
The origin test to determine which side of the line is appropriate only works when
the boundary line itself does not go through the origin. If it does, then some other point
nah77856_ch03Supple.indd 175
17/12/14 11:58 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 176
nah77856
176
Supplement One
Linear Programming
176 Supplement
One Linear
Programming
FIGURE S1–3
Graphing a constraint
boundary that passes
through the origin
4x1 – 3x2 = 0
x2
4
3
2
1
Test point (1, 0)
1
2
3
4
x1
not on the line must be used for the test. For example, consider the constraint
4x1 � 3x2 � 0.
When we try to graph the line 4x1 � 3x2 � 0, we see that substituting x1 � 0 gives x2 � 0
and substituting x2 � 0 gives x1 � 0. This means that the line passes through the origin.
In order to graph the line, we must determine another point that lies on it. Just pick any
value of x1 and solve for the corresponding value of x2. For example, substituting x1 � 3
gives x2 � 4, meaning that the point (3, 4) lies on the line as well as the point (0, 0). The
boundary line is graphed by connecting these points, as shown in Figure S1–3.
Next, we determine which side of the inequality corresponds to the region of interest. As noted above, the origin test does not work when the line passes through the origin. We pick any point not on the line to do the test. In this case, one point that does not
lie on the line is x1 � 1 and x2 � 0. Substituting these values into the inequality gives
(4)(1) � 0 � 4 � 0.
The inequality is satisfied, so the point (1, 0) lies in the region. The inequality corresponds to the points below the line as pictured in Figure S1–3.
Graphing the Feasible Region
The graph of the feasible region is found by graphing the linear inequalities represented by the constraints and determining the region of intersection of the corresponding half spaces. We will determine a graph of the feasible region for the Sidneyville
problem in this way.
We have graphed the feasible region corresponding to the first constraint in Figure S1–2. Consider the other two constraints:
4x1 � 16x2 � 128,
15x1 � 10x2 � 220.
The half spaces corresponding to these constraints are found in the same way. First
we graph the straight lines corresponding to the region boundaries. In the first case
nah77856_ch03Supple.indd 176
17/12/14 11:58 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 177
S1.4 Solving Linear
Linear Programming
Programming Problems
Problems Graphically 177
Graphically 177
S1.4 Solving
FIGURE S1–4
Feasible region for
Sidneyville problem
(Example S1.1)
22
20
15x1 + 10x2 = 220
10
8
4x1 + 16x2 = 128
x2
x1
10
14.6667
20
30 32
10x1 + 20x2 = 200
the intercepts are x1 32 and x2 8, and in the second case they are x1 14.6667 and
x2 22. Using the origin test, we see that the appropriate half spaces are the points
lying below both lines. In addition, we must also include the nonnegativity constraints,
x1 0 and x2 0. The resulting feasible region is pictured in Figure S1–4.
Finding the Optimal Solution
The feasible region pictured in Figure S1–4 has several interesting properties that are
common to all linear programming problems. Pick any two feasible solutions (that is,
points in the region) and connect these points with a straight line. The resulting line lies
completely in the region, meaning that the feasible region is a convex set. The region
boundaries are straight lines. These lines intersect at points known as the extreme points.
There are a total of five extreme points in the feasible region pictured in Figure S1–4.
An important property of linear programs is that the optimal solution always occurs
at an extreme point of the feasible region. For the Sidneyville problem, this means that
among the infinite number of solutions (i.e., points) in the feasible region, the optimal
solution will be one of only five points!2 This is true no matter what objective function
we assume. This means that we can find the optimal solution to the Sidneyville problem by identifying the five extreme points, substituting the (x1, x2) coordinates of these
points into the objective function, and determining the point that results in the maximum profit. We will consider this method first, even though there is a more efficient
graphical solution procedure.
From Figure S1–4 we see that one of the extreme points is the origin (0, 0). Another
is the x2 intercept corresponding to the constraint 4x1 16x2 128, which is (0, 8). A
third is the x1 intercept corresponding to the constraint 15x1 10x2 220, which is
(14.6667, 0). The other two extreme points correspond to the intersections of pairs of
boundary lines. They are found by simultaneously solving the equations corresponding
to the boundaries.
2
It can happen that two extreme points are optimal, in which case all the points on the line connecting
them are optimal as well.
nah77856_ch03Supple.indd 177
17/12/14 11:58 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 178
nah77856
178
Supplement One
Linear Programming
178 Supplement
One Linear
Programming
First, we simultaneously solve the equations
4x1 16x2 128,
10x1 20x2 200.
These equations can be solved in several ways. Multiplying the first equation by 10
and the second by 4 gives
40x1 160x2 1,280,
40x1 80x2 800.
Subtracting the second equation from the first yields
80x2 480
x2 6.
The value of x1 is found by substituting x2 6 into either equation (both will yield
the same value for x1). Substituting into the first equation gives
4x1 (16)(6) 128
4x1 128 96 32
x1 8.
Check that substituting x2 6 into the equation 10x1 20x2 200 gives the same
result.
The last extreme point is found by solving
15x1 10x2 220,
10x1 20x2 200
simultaneously. We will not present the details of this calculation. The reader should be
able to show by the same methods just used that the simultaneous solution in this case is
x1 12,
x2 4.
We have now identified all five extreme points. The next step is to substitute the corresponding values of x1 and x2 into the objective function and see which gives the
largest profit. The objective function is 115x1 90x2.
Extreme Point
Value of Objective Function
(0, 0)
(115)(0) (90)(0) 0
(0, 8)
(115)(0) (90)(8) 720
(14.666 . . . , 0)
(115)(14.666 . . . ) (90)(0) 1,686.67
(8, 6)
(115)(8) (90)(6) 1,460
(12, 4)
(115)(12) (90)(4) 1,740
The maximum value of the objective function is 1,740 and is achieved at the point
(12, 4). This agrees with the computer output in Table S1–1.
Hence, we have shown that one method of finding the optimal solution to a linear
programming problem is to find all the extreme points, substitute their values into the
objective function, and pick the one that gives the largest objective function value
nah77856_ch03Supple.indd 178
17/12/14 11:58 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 179
S1.4 Solving Linear
Linear Programming
Programming Problems
Problems Graphically 179
Graphically 179
S1.4 Solving
for maximization problems or the smallest objective function value for minimization
problems. Next we show how one can quickly identify the optimal extreme point
graphically.
Identifying the Optimal Solution Directly by Graphical Means
One identifies the optimal solution directly in the following way. The objective function is a linear combination of the decision variables. In our example the objective
function is 115x1 � 90x2. Consider the family of straight lines defined by the equation
Z � 115x1 � 90x2.
As Z is varied, one generates a family of parallel lines. The variable Z is the profit obtained when producing x1 rolltop desks and x2 regular desks such that (x1, x2) lies on the
line Z � 115x1 � 90x2. As an example, consider Z � 4,140. Figure S1–5 shows the line
4,140 � 115x1 � 90x2. Notice that it lies completely outside the feasible region, meaning that no feasible combination of x1 and x2 results in a profit of $4,140. Reducing Z to
2,070 drops the Z line closer to the feasible region, as also pictured in Figure S1–5.
The graphical method of identifying the optimal solution is to pick one value of Z,
such as Z � 3,000, that takes us beyond the feasible region, place a ruler on the Z line,
and move the ruler parallel to the Z line toward the feasible region. The extreme point
that is hit first is the optimal solution. Once one graphically determines which extreme
point is optimal, the coordinates of that point are found by solving the appropriate
equations as shown above. This approach avoids having to identify all the extreme
points of the feasible region.
One problem can arise. If we pick a small starting value of Z and move toward the
feasible region from below, this method will identify a different extreme point. In Figure S1–5, suppose that we chose Z � �3,000. Then the Z line would be located under
the feasible region. As we moved the Z line toward the feasible region, the first extreme
point encountered is the origin. This means that the origin solves Example S1.1 with a
FIGURE S1–5
Approaching the
feasible region with
the Z line
x2
46
40
30
20
10
10
20
30
40
Z line
Z = 2,070
Z = 4,140 x1
from below
nah77856_ch03Supple.indd 179
17/12/14 11:58 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 180
nah77856
180
Supplement One
Linear Programming
180 Supplement
One Linear
Programming
minimization objective rather than a maximization objective (i.e., if your goal is to
minimize profit, the best strategy is not to produce any desks).
Hence, this approach identifies two extreme points. One corresponds to the maximum solution and one to the minimum solution. In this case it is obvious which is
which. When it is not obvious, the (x1, x2) values for each extreme point should be
found and substituted into the objective function to be certain which is the minimum
solution and which is the maximum solution.
S1.5
THE SIMPLEX METHOD: AN OVERVIEW
The Simplex Method is an algorithm that moves sequentially from one extreme point
to another until it reaches the optimal solution. If the origin (that is, all problem variables set to zero) is a feasible solution, it will serve as the initial extreme point.3 At each
iteration, the method considers all adjacent extreme points (those that can be reached
by moving along an edge), and moves to the one that gives the best improvement in the
objective function. The algorithm continues to move from one adjacent extreme point
to another, finally terminating when the optimal solution is reached.
For the problem of Example S1.1 the origin is feasible, so that is the initial feasible
solution. The two extreme points adjacent to the origin are (x1, x2) (14.6667, 0) and
(x1, x2) (0, 8). The largest improvement in the objective function is obtained by moving to the point (14.6667, 0). There are two extreme points adjacent to (14.6667, 0).
They are (0, 0) and (12, 4). Clearly, a greater improvement is obtained by moving to
(12, 4). At this point the method recognizes that the current solution is optimal, as a
movement to another adjacent extreme point lowers the profit.
In the worst case, the Simplex Method could conceivably need to search all the
extreme points of the feasible region before identifying the optimal solution. If this
were common, the Simplex Method would not be a practical solution method for solving linear programming problems. Let us consider why.
Suppose that a linear program had 25 variables and 25 less than or equal to constraints. For each constraint, one adds a slack variable converting the problem to one
with only equality constraints. This is known as standard form. Hence, in standard
form we have a problem of 50 variables and 25 constraints. Each basic solution
(extreme point) corresponds to setting 25 variables to zero and solving the resulting
system of 25 linear equations in 25 unknowns. It follows that the number of such
solutions (that is, extreme points) equals the number of combinations of 50 things
taken 25 at a time. This turns out to be about 1.264 1014 (about 126 trillion). To
give the reader some idea of the magnitude of this number, suppose that we had
a computer program that could identify 100 extreme points every second. At that
rate, it would take about 40,000 years to find all the extreme points for a problem of
this size!
It is indeed fortunate that the Simplex Method rarely needs to search all the extreme
points of the feasible region to discover the optimal solution. In fact, for a problem of
this size, on average it would need to evaluate only about 25 extreme points.4 Hence, the
Simplex Method turned out to be a very efficient solution procedure.
3
When the origin is not feasible, there are techniques available for determining an initial feasible solution
to get the method started.
4
The reason for this was understood only very recently. The proof requires very sophisticated
mathematics.
nah77856_ch03Supple.indd 180
17/12/14 11:58 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 181
S1.6 Solving Linear
Linear Programming
Programming Problems
Problems with
with Excel 181
Excel 181
S1.6 Solving
We will not explore the mechanics of the method or additional theory underpinning
its concepts. With today’s easy access to computing and the wide availability of excellent software, it is unlikely that anyone with a real problem would solve it manually.
There are many excellent texts that delve more deeply into the theoretical and computational aspects of linear programming. A good starting point for the interested reader
is Hillier and Lieberman (1990). A more detailed treatment of the theory can be found
in Hadley (1962).
S1.6
SOLVING LINEAR PROGRAMMING PROBLEMS WITH EXCEL
Excel has become the standard for spreadsheet programs. One of the useful features of
Excel is that it comes bundled with an add-in called Solver that solves both linear and
nonlinear programming problems. While the reliability of the nonlinear and the integer
portions of the program are suspect, the linear programming part of Solver is excellent.
Because Solver is part of a spreadsheet program, problems are not entered algebraically (as they are with a system known as LINDO discussed in previous editions of
this book). Consider Example S1.1. The algebraic representation is
Maximize 115x1 � 90x2,
subject to
10x1 � 20x2 � 200,
4x1 � 16x2 � 128,
15x1 � 10x2 � 220,
x1, x2 � 0.
It is convenient to write the problem in a matrix format before entering the information
into the Excel spreadsheet. The matrix representation for this problem is
Variable names:
Objective function:
subject to
Constraint 1
Constraint 2
Constraint 3
x1
115
x2
90
Operator
max
RHS
10
4
15
20
16
10
�
�
�
200
128
220
The spreadsheet will look very much like this. The only differences are that one must
specify the locations of the variable values (which I recommend be a row located directly
under the variable names) and the algebraic formulas for the objective function and the
constraints. These will be located in a column between “Operator” and “RHS.” We will
label this column “Value” in the spreadsheet.
Note that the column labeled “Operator” is not required. It is a convenience for the
user to help keep track of the direction of the objective function and constraints. Excel
requires this information, and it is entered manually when Solver is invoked.
At this point the spreadsheet should look similar to Figure S1–6. Notice the additional row labeled “Var Values” and the additional column labeled “Value.” It is in the
column labeled “Value” that one enters the algebraic form of the linear programming
nah77856_ch03Supple.indd 181
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 182
nah77856
182
Supplement One
Linear Programming
182 Supplement
One Linear
Programming
FIGURE S1–6
Excel spreadsheet for Sidneyville problem
problem. The locations of the variable values are cells C4 and D4. The algebraic form
of the objective function will be entered in cell F5 and the constraints in cells F7, F8,
and F9. The formula for cell F5 is C5*$C$4D5*$D$4. The formula can be typed
in or entered using the mouse to point and click cell locations.
Notice that we have used absolute addressing for the variable values (C4 and D4).
This allows us to copy the formula from cell F5 to cells F7, F8, and F9 without having
to retype the algebraic form for the constraints. You may wish to assign name labels to
these cells so that they can later be referred to by name rather than by cell location.
(This is done most conveniently by invoking the formula bar, accessing the label area,
and typing in a label of your choice. Note that the label “profit” appears just below the
name of the current font. This was the name assigned to cell F5.)
Check that the algebraic form of the constraints is correct after copying cell F5 to
cells F7, F8, and F9. For example, the formula corresponding to cell F7 should be:
C7*$C$4 D7*$D$4.
Place 0s in the cells corresponding to the variable values (C4 and C5). After Solver
has completed its search, the optimal values of these variables will appear in these cell
locations.
The problem is now completely defined. Invoke Solver to obtain the solution. This
is done by accessing Tools from the Menu Bar and choosing Solver. The Solver dialog
nah77856_ch03Supple.indd 182
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 183
S1.6 Solving Linear
Linear Programming
Programming Problems
Problems with
with Excel 183
Excel 183
S1.6 Solving
FIGURE S1–7
Spreadsheet with Solver dialog box
box will appear as in Figure S1–7. The first requirement is setting the location of the
target cell. This corresponds to the cell location of the algebraic formula for the objective function—cell F5 in our spreadsheet. This can be accomplished by typing in F5 or
clicking on the cell with the mouse. Notice that in this case I assigned the name “profit”
to cell F5, so I simply type in “profit” for the target cell.
Next, specify if the problem is a min (minimum) or a max (maximum). Excel refers
to the variables in the problem as “changing cells.” You must tell Solver where to find
these cells. In this spreadsheet they are C4 and D4. (These can be indicated by pointing and clicking with the mouse, manually typing in the cell locations, or entering
preassigned cell names as we did for the objective function.)
Next, tell Excel where to find the algebraic definitions for the constraints. The
constraints are entered into the system one at a time by clicking on the Add button. For
each constraint you first tell the system where to find the algebraic form for the lefthand side of the constraint (F7, F8, and F9 in our case), the logical operator for the constraint (, , or ), and the location of the RHS value (G7, G8, and G9 in our case).
Because Solver is a general-purpose mathematical programming tool, two additional pieces of information must be included. Both of these can be done at the same
time by clicking the Options key and choosing the two options “Assume linear model”
and “Assume non-negative.” In this way, the nonnegativity constraints do not need
nah77856_ch03Supple.indd 183
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 184
184
Supplement One
Linear Programming
184 Supplement
One Linear
Programming
FIGURE S1–8
The Excel spreadsheet displaying the optimal solution
to be explicitly entered. (Note: The nonnegativity option was not available in Excel
Version 5 or earlier. Nonnegativity constraints had to be entered into the problem
explicitly.) Telling the program to assume a linear model ensures that the problem is
solved by the Simplex algorithm rather than the gradient method used for nonlinear
problems. If your output includes values of Lagrange variables, you’ll know that you
forgot to specify this option.
At this point your dialog box should look like Figure S1–7. Notice that I have named
some of the cells and the names appear in the dialog box rather than the cell locations.
Using named cells that have some meaning relative to your problem will be very helpful later when you obtain the solution and sensitivity reports. In the dialog box, we have
named the objective cell (F5) “profit” and the cells corresponding to the constraints
“cons1,” “cons2,” and “cons3.”
Check that all information is correctly entered and that you have specified a linear
problem. Now simply click the mouse on the Solve button and Excel will whirl away
and quickly produce a solution. After solving, the resulting spreadsheet should look
like the one in Figure S1–8. Notice that the values of the variables in cells C4 and D4
now reflect the optimal solution of 12 and 4, respectively. The value in cell F5 is the
value of the optimal profit of $1,740. The values in cells C7, C8, and C9 are the values
of the left-hand sides of the constraints.
nah77856_ch03Supple.indd 184
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 185
S1.6 Solving Linear
Linear Programming
Programming Problems
Problems with
with Excel 185
Excel 185
S1.6 Solving
FIGURE S1–9
Answer report for Sidneyville problem
Although the optimal values of the variables and the optimal value of the objective
function now appear in your spreadsheet, Excel has the option of printing several types of
reports. The two that are relevant for linear programming are the Answer and Sensitivity
reports. These reports appear on different sheets of the Excel workbook and are shown
in Figures S1–9 and S1–10. Most of the information in the Answer report appears in the
original spreadsheet. We also are told which constraints are binding. A variable that is
labeled “not binding” is one that is nonzero.
Entering Large Problems Efficiently
The steps outlined for solving linear programming problems on a spreadsheet are fine
for small problems. Excel has several features that allow more efficient entry of larger
problems, however.
One such feature is the SUMPRODUCT function. SUMPRODUCT is a vector or
array product, which multiplies the elements of one array times another array term
by term and adds the results. This means that the algebraic form for the objective
function and for the constraints can be entered with this function. Recall that for cell
F5 we used the formula C5*$C$4D5*$D$4. This formula also could have been
entered as SUMPRODUCT($C$4:$D$4, C5:D5). While this may not appear to be
much of an improvement here, it saves a lot of typing when entering large problems.
nah77856_ch03Supple.indd 185
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 186
186
Supplement One
Linear Programming
186 Supplement
One Linear
Programming
FIGURE S1–10
Sensitivity report for Sidneyville problem
The second shortcut for entering large problems is to group constraints by type and
enter all constraints in a group at one time. In the case of the Sidneyville problem, there
are three constraints. These constraints could be entered with one command, similar to
the single command for entering the nonnegativity constraints. The appropriate formulas
for the constraints appear in cells F7, F8, and F9. The single command for entering these
three constraints into Solver is: $F$7:$F$9 $G$7:$G$9. This can be typed in directly
or entered by pointing to the appropriate cells in the spreadsheet. The Solver dialog box
for the Sidneyville problem using this approach is shown in Figure S1–11.
Using SUMPRODUCT and entering constraints in groups can save a lot of time for
large problems. However, the Solver dialog box is not as informative, since only cell
locations and not cell names appear. For problems of fewer than 10 constraints total,
entering the constraints one at a time is fine.
We noted earlier that an advantage of the spreadsheet for solving linear programming is that one can construct a general-purpose template. A template might have up
to 10 variables and 10 constraints. Variable names could be x1 through x10 and
constraint names const1 through const10. The SUMPRODUCT functions for the
objective function and the constraints would be programmed in advance. One would
then simply enter the coefficients of the problem to be solved and save as a new file
name. (This is much faster than typing in the full algebraic representation for every
new problem as one must do with other linear programming systems.)
nah77856_ch03Supple.indd 186
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 187
S1.7 Interpreting the
the Sensitivity
Sensitivity Report 187
Report 187
S1.7 Interpreting
FIGURE S1–11
Solver dialog box with efficient data entry
S1.7
INTERPRETING THE SENSITIVITY REPORT
Shadow Prices
An interesting issue is the value of additional resources in a linear programming
problem. In Example S1.1, the resources are the three types of wood: pine, cedar, and
maple. An increase in the level of any of these resources results in an increase in
the value of the right-hand side of the appropriate constraint. The sensitivity report
(Figure S1–10) gives information about the value to the objective function of additional resources. This information is contained in the values of the shadow prices for
the constraints.
The shadow price is defined as the improvement in the objective function realized
by adding one additional unit of a resource. For Example S1.1, the first constraint
refers to the amount of pine needed. Since this constraint is binding at the optimal
solution (since the final value and the constraint right-hand-side values are the same),
it is likely that if we had additional pine we could increase revenue. The shadow price
tells us just how beneficial. Since the shadow price for this constraint is $1, it means
that for each additional unit square foot of pine, the objective (profit) increases by $1.
(This will hold only within a certain range as discussed subsequently.) Consider the
nah77856_ch03Supple.indd 187
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 188
nah77856
188
Supplement One
Linear Programming
188 Supplement
One Linear
Programming
second constraint (cedar). Since the final value and the right-hand side for this constraint are different (112 versus 120), there is slack in this constraint. That means we
are not consuming the entire quantity of cedar at the optimal solution, and additional
cedar will not improve the profit. This is borne out by the fact that the shadow price for
this constraint is zero. The final shadow price of 7 indicates that every additional
square foot of maple contributes $7 to the profit.
Objective Function Coefficients and Right-Hand Sides
The shadow prices remain valid as long as the optimal basis does not change. (The
basis is the set of positive variables in the final solution.) The columns Allowable
Increase and Allowable Decrease indicate over what range the shadow prices remain
valid. This means that we can determine the effect on the objective function of changes
in constraint right-hand sides without re-solving the problem.
The first part of the sensitivity report (Figure S1–10) gives the values of the objective function coefficients for which the shadow prices remain valid. The current values
of the profits are $115 and $90, respectively. The shadow prices are valid as long as the
first objective function coefficient does not increase more than 20 or decrease more
than 70 (i.e., the objective function coefficient for x1 is between 45 and 135). Similarly,
the allowable range for the objective function coefficient for x2 is 76 23 to 230.
The second part of the sensitivity report reports the ranges on the right-hand sides of
the constraints for which the shadow prices remain valid. Hence, the shadow price of $1
for pine is valid for an increase of 16 or less and a decrease of 53 13 or less in the righthand side of the first constraint. That is, the right-hand side of the first constraint could
be any value between 146 23 and 216. The shadow price of 0 for cedar is valid for any increase (1E30 should be interpreted as infinity) and a decrease of 16 or less, and the
shadow price for maple is valid for an increase of 80 or less and a decrease of 40 or less.
If either the objective function coefficients or the right-hand sides increase or decrease beyond the allowable ranges, the shadow prices no longer remain valid, and the
problem would have to be re-solved to determine the effect. Note that Excel uses the
convention that a positive shadow price means an increase in the objective function per
unit increase in the right-hand side of a constraint, and a negative shadow price means
a decrease in the objective function per unit increase in the right-hand side, irrespective of whether the problem is a max or a min. (Other linear programming systems may
have other conventions.) Also note that changes to objective function coefficients or
right-hand sides can be only one at a time in these ranges. The rules for simultaneous
changes in right-hand sides or objective function coefficients are much more complex
and will not be discussed here.
Adding a New Variable
We can use the results of sensitivity analysis to determine whether it is profitable to add
a new activity (variable) without re-solving the problem. For small problems, such as
Example S1.1, simply inputting and solving the new problem on the computer is quick
and easy. However, in real applications, the number of decision variables and constraints could be in the hundreds or even in the thousands. Reentering a problem of this
magnitude is a major task, to be avoided whenever possible.
Suppose the firm is considering producing a third product, a vanity table, which
would require the same woods used in making the desks. Each vanity table would contribute $75 to profit, but each would require 8 square feet of pine, 6 square feet of
cedar, and 10 square feet of maple. We could determine if it would be worth producing
vanity tables in addition to the desks by solving the problem with three activities and
comparing the values of the objective functions at the optimal solutions.
nah77856_ch03Supple.indd 188
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 189
S1.7 Interpreting the
the Sensitivity
Sensitivity Report 189
Report 189
S1.7 Interpreting
There is a faster way. The dual prices in Figure S1–10 tell us the value of each unit
of resource at the current solution. The decrease in profit resulting from reducing the
supply of pine is $1 per square foot, which translates to $8.00 for 8 square feet of pine.
There is no cost for decreasing the supply of cedar. The cost of decreasing the supply
of maple by 10 square feet is (10)(7) $70. Hence the total decrease in profit from the
consumption of resources required to produce one vanity table is $78. The contribution
to profit is only $75. We conclude that it is not optimal to produce vanity tables in
addition to the desks with the current resources. Had we determined, however, that it
was profitable to produce the vanity table, we would have had to re-solve the problem
with three activities to find the optimal numbers of desks and vanity tables to produce.
Using Sensitivity Analysis
To cement your understanding of the information in Figure S1–10, consider the following questions:
Example S1.1
(continued)
a. Sidneyville’s sales manager has renegotiated the contract for regular desks and now expects
to make a profit of $125 on each. He excitedly conveys this information to the firm’s production manager, expecting that the optimal mix of rolltop and regular desks will change as
a result. Does it?
b. Suppose that the new contract also has a higher profit for the rolltop desks. If the new profit
for the rolltop desks is $140, how will this change the optimal solution?
c. A logging company has offered to sell Sidneyville an additional 50 square feet of maple for
$5.00 per square foot. Based on the original objective function, would you recommend that
it accept the offer?
d. Assuming that Sidneyville purchases the 50 square feet of maple, how is the optimal solution
affected?
e. The firm is considering a pine desk that would require 25 square feet of pine and no other
wood. What profit for pine desks would be required to make its production worthwhile,
assuming current levels of resources and original profits on regular and rolltop desks?
f. During inspection, the quality department discovered that 50 square feet of pine had water
damage and could not be used. Will it be optimal to produce both desks under these
circumstances? Will the product mix change?
Solution
a. According to Figure S1–10, the allowable increase in the coefficient of the objective function for variable x2, the regular desks, is 140. Because the increase to $125 is still within the
allowable range, the optimal mix of rolltop and regular desks will remain the same: namely,
x1 12 and x2 4.
b. The allowable increase in the objective function for the rolltop desks (x1) is 20, or to a maximum value of 135. As 140 is outside the allowable range, it is possible that the basis will
change. However, the allowable ranges in Figure S1–10 are only valid if the profit for regular desks is $90. The allowable ranges will change when the profit for regular desks is
changed to $125, even though the optimal solution does not. The output for part (a) (that is,
with profits of $115 and $125) is
OBJ COEFFICIENT RANGES
VARIABLE
X1
X2
nah77856_ch03Supple.indd 189
CURRENT
COEF
115.000000
125.000000
ALLOWABLE
INCREASE
72.500000
105.000000
ALLOWABLE
DECREASE
52.500000
48.333330
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 190
nah77856
190
Supplement One
Linear Programming
190 Supplement
One Linear
Programming
This shows that the allowable increase in the coefficient for x1 is now 72.5. Because 140 is
within the allowable range, the solution for parts (a) and (b) will be the same, which is also the
same as our original solution of x1 12 and x2 4.
c. Since the dual price for the third constraint corresponding to maple is 7, it is profitable to
purchase the maple for $5.00 a square foot. The allowable increase of the right-hand side
over which this dual price applies is 80, so it is worth purchasing the full 50 additional
square feet.
d. Because the increase of 50 is within the allowable right-hand-side range, we know that the
basis will not change. That is, it still will be optimal to produce both the rolltop and the regular desks. However, if the right-hand side changes, the values of the basic variables will
change. We must re-solve the problem with the new right-hand-side value to determine the
updated solution. The solution is
LP OPTIMUM FOUND AT STEP 1
OBJECTIVE FUNCTION VALUE
1)
VARIABLE
X1
X2
2090.000000
VALUE
17.000000
1.500000
REDUCED COST
.000000
.000000
To retain feasibility we round x2 to 1. (This is not the optimal integer solution, however.
The optimal integer solution is x1 18 and x2 0 with a profit of $2,070, which is obtained
from Excel by identifying both x1 and x2 as integer variables. The suboptimal solution of
x1 17 and x2 1 results in a profit of $2,045.)
e. The dual price for pine is $1 per square foot. As each desk consumes 25 square feet of
pine, the profit for each pine desk must exceed $25 for pine desks to be profitable to
produce.
f. The right-hand side of the first constraint can decrease as much as 53.333330 and the
current basis will remain optimal. That means that a decrease of 50 square feet will not
change the basis; it will still be profitable to produce both desks. However, the production
quantities will decrease. We must re-solve the problem to determine the correct levels of
the new quantities. They are
LP OPTIMUM FOUND AT STEP 2
OBJECTIVE FUNCTION VALUE
1)
1690.000000
VARIABLE
VALUE
X1
14.500000
X2
.250000
REDUCED COST
.000000
.000000
Again, we need to round these variables. Rounding both x1 and x2 down guarantees feasibility. If we produce 14 rolltop desks and 0 regular desks, we require 140 square feet of
pine (150 are available), 56 square feet of cedar (128 are available), and 210 square feet of
maple (220 are available). There does not appear to be enough wood to produce an additional desk of either type, so we leave the solution at x1 14 and x2 0. (This is the optimal
integer solution.)
nah77856_ch03Supple.indd 190
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 191
S1.8 Recognizing Special
Special Problems 191
Problems 191
S1.8 Recognizing
S1.8
RECOGNIZING SPECIAL PROBLEMS
Several problems can occur when solving linear programming problems. In this section we will discuss the causes of these problems and how one recognizes them when
using Excel.
Unbounded Solutions
The feasible region of a linear program is not necessarily bounded. The feasible region
for the Sidneyville problem pictured in Figure S1–4 is bounded. However, consider the
following linear programming problem.
Example S1.2
Maximize
2x1 3x2
subject to
x1 4x2 8,
x1 x2 5,
2x1 x2 7,
x1, x2 0.
Figure S1–12 shows the feasible region. Notice that it is unbounded. Because we can make
x1 and x2 as large as we like, there is no limit to the size of the objective function. When this
occurs, the problem is unbounded and there is no optimal solution.
When this problem is inputted in Excel, Solver writes in very large values for the problem
variables and displays a message that set target values do not converge. The Excel output for
this problem appears in Figure S1–13.
FIGURE S1–12
Feasible region for
Example S1.2
7
6
5
x2
4
3
2
1
1
2
2x1 + x2 = 7
3
4
5
6
x 1 + x2 = 5
7
8
x1 + 4x2 = 8
x1
nah77856_ch03Supple.indd 191
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 192
nah77856
192
Supplement One
Linear Programming
192 Supplement
One Linear
Programming
FIGURE S1–13
Excel output for Example S1.2
Empty Feasible Region
It is possible for two or more constraints to be inconsistent. When that occurs, there
will be no feasible solution. Consider the following problem.
Example S1.3
Maximize
2x1 3x2
subject to
x1 4x2 8,
x1 x2 10,
x1, x2 0.
The feasible region for this example appears in Figure S1–14. Notice that there is no intersection of the half spaces defined by the two constraints in the positive quadrant. In this case,
the feasible region is empty and we say that the problem is infeasible. The Excel output appears
in Figure S1–15. Note that the solution x1 8 and x2 0 shown is not feasible because it
results in negative slack in the first constraint.
nah77856_ch03Supple.indd 192
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 193
S1.8 Recognizing Special
Special Problems 193
Problems 193
S1.8 Recognizing
FIGURE S1–14
Feasible region for
Example S1.3
10
9
8
7
6
x2 5
4
3
2
x1 + x2 = 10
x1 + 4x2 = 8
1
1
2
3
4
5
6
7
8
9
10
x1
FIGURE S1–15
Excel output for Example S1.3
nah77856_ch03Supple.indd 193
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 194
nah77856
194
Supplement One
Linear Programming
194 Supplement
One Linear
Programming
Degeneracy
In linear programming there are two types of variables: basic variables and nonbasic variables. The number of basic variables equals the number of constraints, and basic variables
may be either original variables or slack or surplus variables. What defines basic variables? Consider any linear program in standard form. By including slack and surplus
variables, all constraints are expressed as equations. In standard form, there always will
be more variables than constraints. Suppose that after a linear programming problem has
been expressed in standard form, there are n m variables and n constraints. A basic solution is found by setting m variables to zero and solving the resulting n equations in n unknowns. In most cases the values of the n basic variables will be positive. A degenerate
solution occurs when one or more basic variables are zero at the optimal solution. In
Excel, a basic variable is one with zero reduced cost. Degeneracy occurs when the value
of a variable is zero and its reduced cost or shadow price is also zero.
Why are we interested in degeneracy? It is possible that if degenerate solutions
occur, the Simplex Method will cycle through some set of solutions and never recognize the optimal solution. The phenomenon of cycling has never been observed in
practice, and most computer programs have means of guaranteeing that it never occurs.
The bottom line is that degeneracy is an issue about which we need not worry, but one
of which we should be aware.
Multiple Optimal Solutions
The optimal solution to a linear program is not always unique. There are cases in which
there are multiple optimal solutions. In Chapter 3 we saw that two-variable problems
could be solved by graphical means by approaching the feasible region with the Z line.
Assuming that we approach the feasible region from the correct side, the first feasible
point with which the Z line comes into contact is the optimal solution.
However, suppose that the Z line is parallel to one of the constraints. In that case it
does not contact a single point first, but an entire edge.
Example S1.4
Consider the feasible region pictured in Figure S1–12 corresponding to the constraints of
Example S1.2. Suppose that the objective function is min 3x1 3x2. Then the Z line has slope
1 and is parallel to the constraint boundary x1 x2 5. As the Z line approaches the feasible region, it meets the edge defined by this constraint and both extreme points along this
edge. This means that both extreme points and all points along the edge are optimal.
Unfortunately, Excel does not indicate that there are multiple optimal solutions to this problem.
Our only clue is that the solution is degenerate; the surplus variable for the third constraint has
both zero dual price and zero value. Our graphical solution tells us that both extreme points (2, 3)
and (4, 1) are optimal, and so are all points along the edge connecting these extreme points. [The
points along the edge can be written in the form
x1 (2) (1 )(4),
x2 (3) (1 )(1),
where is a number between zero and one. This is known as a convex combination of these
two extreme points.]
Redundant Constraints
It is possible for one or more constraints to be redundant. That means that these
constraints can be eliminated from the formulation without affecting the solution. In
simple two-variable problems, redundant constraints can be recognized graphically
because they lie outside the feasible region. Excel does not recognize or signal if one
or more constraints are redundant. Sometimes redundant constraints can cause degeneracy, but degeneracy can result when constraints are not redundant as well.
nah77856_ch03Supple.indd 194
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 195
S1.9 The Application
Application of
of Linear
Linear Programming
Programming to
to Production
Production and
and Operations
Operations Analysis 195
Analysis 195
S1.9 The
FIGURE S1–16
Feasible region
for Example S1.5
showing a redundant
constraint
20
10
x2
x1
Example S1.5
10
20
x1 + x2 20
Consider Example S1.1. Suppose that we add the following additional constraint:
x1 x2 20.
Figure S1–16 shows the resulting feasible region. It is exactly the same as the feasible region
pictured in Figure S1–4. The additional constraint has no effect, as it lies completely outside
the feasible region defined by the first three constraints. The optimal solution will, of course, be
exactly the same.
If we had originally formulated the problem with the four constraints
10x1 20x2 200,
4x1 16x2 128,
15x1 10x2 220,
x1 x2 20,
and solved, the output indicates that the optimal solution is x1 12 and x2 4 as before, and
gives us no clue that the last constraint is redundant. The only way to see that the final constraint is redundant is to graph the feasible region as we did in Figure S1–16. Graphing is possible, however, in two-variable problems only.
Does redundancy cause a problem? Not really. We would certainly like to be able to write
our linear program as economically as possible, but if one or more constraints are redundant,
the optimal solution is unaffected.
S1.9 THE APPLICATION OF LINEAR PROGRAMMING
TO PRODUCTION AND OPERATIONS ANALYSIS
In Chapter 3 we showed how linear programming could be used to find optimal solutions (subject to rounding errors) for aggregate planning problems. Although this is the
only explicit use of linear programming in this book, there have been successful linear
nah77856_ch03Supple.indd 195
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 196
nah77856
196
Supplement One
Linear Programming
196 Supplement
One Linear
Programming
programming applications for many operations management problems.5 Scheduling
and distribution are perhaps two areas in which applications are most common.
Fisher et al. (1982) describe an application of linear programming to the problem
of providing a coordinated vehicle scheduling and routing system for delivery of
consumable products to customers of the Du Pont Company. The primary issue was to
determine delivery routes (loops) for the trucks to the company’s clients in various
regions of the country. A refrigerated truck drives a weekly loop that includes several
dozen customers. The largest region considered, Chicago, had 16 loops and several
hundred cities, whereas the Houston region, the smallest, had 4 loops and less than
80 cities.
The basic mathematical formulation of the problem was a generalized assignment
problem. (The assignment problem is discussed in Chapter 9. It is a linear programming
problem in which the decision variables are restricted to be zeros and ones.) The mathematical formulation used in this study is the following:
1. Given data
dik Cost of including customer i in loop k,
ai Demand from customer i.
2. Problem variables
yik e
1 if customer i is assigned to loop k,
0 if customer i is not assigned to loop k.
3. Generalized assignment problem
Min a a dik yik
K
n
k1 i1
subject to
a yik 1,
K
a ai yik bk,
n
for i 1, . . . , n,
k1
for k 1, . . . , K,
i1
yik 0 or 1, for all i and k,
where K is the total number of loops in the region and n is the number of customers.
The implementation of this model was reported to have saved Du Pont over $200 million. A more complex mathematical model solving a similar problem for Air Products
Corporation was reported by Bell et al. (1983). This study won the Institute of Management Sciences Practice Award in 1983.
Linear programming (or, more generally, mathematical programming) has been an
important tool for logistics planning for a wide variety of operations management
problems. Today, Microsoft bundles Solver, a general purpose mathematical programming Excel add-in, with its office software, thus making linear programming accessible to a much wider audience.
5
nah77856_ch03Supple.indd 196
In this section we interpret linear programming in the broad sense to include integer linear programming.
17/12/14 11:59 AM
nah77856_ch03Supple.qxd
1/8/08
3:01 PM
Page 197
Bibliography 197
Bibliography 197
Bibliography
Bell, W. J.; L. M. Dalberto; M. L. Fisher; A. J. Greenfield; R.
Jaikumar; P. Kedia; R. G. Mack; and P. J. Prvtzman.
“Improving the Distribution of Industrial Gases with an
On-Line Computerized Routing and Scheduling Optimizer.”
Interfaces 13 (1983), pp. 4–23.
Fisher, M.; A. J. Greenfield; R. Jaikumar; and J. T. Uster III.
“A Computerized Vehicle Routing Application.” Interfaces
12 (1982), pp. 42–52.
nah77856_ch03Supple.indd 197
Hadley, G. Linear Programming. Reading, MA: AddisonWesley, 1962.
Hillier, F. S., and G. J. Lieberman. Introduction to
Operations Research. 5th ed. San Francisco: Holden Day,
1990.
17/12/14 11:59 AM
Chapter Four
Inventory Control
­Subject to Known
­Demand
“We want to turn our inventory faster than our people.”
—James Sinegal
Chapter Overview
Purpose
To consider methods for controlling individual item inventories when product demand
is assumed to follow a known pattern (that is, demand forecast error is zero).
Key Points
1. Classification of inventories
• Raw materials. These are resources required for production or processing.
• Components. These could be raw materials or subassemblies that will later be
included into a final product.
• Work-in-process (WIP). These are inventories that are in the plant waiting for
processing.
• Finished goods. These are items that have completed the production process
and are waiting to be shipped out.
2. Why hold inventory?
• Economies of scale. It is probably cheaper to order or produce in large batches
than in small batches.
• Uncertainties. Demand uncertainty, lead time uncertainty, and supply uncertainty all provide reasons for holding inventory.
• Speculation. Inventories may be held in anticipation of a rise in their value or cost.
• Transportation. Refers to pipeline inventories that are in transit from one location to another.
• Smoothing. As noted in Chapter 3, inventories provide a means of smoothing
out an irregular demand pattern.
• Logistics. System constraints that may require holding inventories.
• Control costs. Holding inventory can lower the costs necessary to monitor a
system. (For example, it may be less expensive to order yearly and hold the
units than to order weekly and closely monitor orders and deliveries.)
198
nah77856_ch04.indd 198
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 199
Chapter Four Four Inventory
Inventory Control
Control S­Subject
to Known
Known ­D
Demand
199
Chapter
ubject to
emand 199
3. Characteristics of inventory systems
• Patterns of demand. The two patterns are (a) constant versus variable and
(b) known versus uncertain.
• Replenishment lead times. The time between placement of an order
(or initiation of production) until the order arrives (or is completed).
• Review times. The points in time that current inventory levels are checked.
• Treatment of excess demand. When demand exceeds supply, excess demand
may be either backlogged or lost.
4. Relevant costs
• Holding costs. These include the opportunity cost of lost investment revenue,
physical storage costs, insurance, breakage and pilferage, and obsolescence.
• Order costs. These generally consist of two components: a fixed component
and a variable component. The fixed component is incurred whenever a
positive order is placed (or a production run is initiated), and the variable
component is a unit cost paid for each unit ordered or produced.
• Penalty costs. These are incurred when demand exceeds supply. In this case
excess demand may be back-ordered (to be filled at a later time) or lost. Lost
demand results in lost profit, and back orders require record keeping and in
both cases, one risks losing customer goodwill.
5. The basic EOQ model. The EOQ model dates back to 1915 and forms the basis
for all the inventory control models developed subsequently. It treats the basic
trade-off between the fixed cost of ordering and the variable cost of holding. If
h represents the holding cost per unit time and K the fixed cost of setup, then we
show that the order quantity that minimizes costs per unit time is Q � 2Kh
,
where is the rate of demand. This formula is very robust for several reasons:
(a) It is a very accurate approximation for the optimal order quantity when
demand is uncertain (treated in Chapter 5), and (b) we show that deviations from
the optimal Q generally result in modest cost errors. For example, a 25 percent
error in Q results in an average annual holding and setup cost error of only
2.5 percent.
6. The EOQ with finite production rate. This is an extension of the basic EOQ model
to take into account that when items are produced internally rather than ordered
from an outside supplier, the rate of production is finite rather than infinite, as
would be required in the simple EOQ model. We show that the optimal size of a
production run now follows the formula Q � 2Kh
� where h� � h(1 � P)
and P is the rate of production (P � ). Note that since h� � h, the batch size
when the production rate is taken into account exceeds the batch size obtained
by the EOQ formula.
7. Quantity discounts. We consider two types of quantity discounts: all-units and
incremental discounts. In the case of all-units discounts, the discount is applied to
all the units in the order, while in the case of incremental discounts, the discount
is applied to only the units above the break point. The all-units case is by far the
most common in practice, but one does encounter incremental discounts in
industry. In the case of all-units discounts, the optimization procedure requires
searching for the lowest point on a broken annual cost curve. In the incremental
discounts case, the annual cost curve is continuous, but has discontinuous
derivatives.
nah77856_ch04.indd 199
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 200
200 Chapter Four
Inventory Control
200 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
8. Resource-constrained multiple product systems. Consider a retail store that orders
many different items, but cannot exceed a fixed budget. If we optimize the order
quantity of each item separately, then each item should be ordered according to
its EOQ value. However, suppose doing so exceeds the budget. In this section, a
model is developed that explicitly takes into account the budget constraint and
adjusts the EOQ values accordingly. In most cases, the optimal solution subject
to the budget constraint requires an iterative search of the Lagrange multiplier.
However, when the condition c1h1 � c2h2 � … � cnhn is met, the optimal
order quantities are a simple scaling of the optimal EOQ values. Note that this
problem is mathematically identical to one in which the constraint is on available
space rather than available budget.
9. EOQ models for production planning. Suppose that n distinct products are
produced on a single production line or machine. Assume we know the holding
costs, order costs, demand rates, and production rates for each of the items. The
goal is to determine the optimal sequence to produce the items, and the optimal
batch size for each of the items to meet the demand and minimize costs. Note
that simply setting a batch size for each item equal to its EOQ value (that is,
optimal lot size with a finite production rate), is likely to be suboptimal since it is
likely to result in stock-outs. The problem is handled by considering the optimal
cycle time, T, where we assume we produce exactly one lot of each item each
cycle. The optimal size of the production run for item j is simply Qj � jT, where
T is the optimal cycle time. Finding T is nontrivial, however.
The current investment in inventories in the United States is enormous. In the third
quarter of 2007, the total dollar investment was estimated to be $1.57 trillion.1 Figure 4–1 shows investment in inventories broken down by sectors of the economy. The
inventory models we will be discussing in this chapter and in Chapter 5 can be applied
FIGURE 4–1
Breakdown of the total
investment in
inventories in the
U.S. economy
(2007)
Other
(11%)
Farm
(7%)
$174.6
billion
$114.6
billion
Retail
(30%)
Manufacturing
(29%)
$472.5
billion
$454.9
billion
Wholesale
(23%)
$353.6
billion
Total investment = $1,570.2 billion
1
nah77856_ch04.indd 200
Survey of Current Business (July 2007).
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 201
4.1 Types of
of Inventories 201
Inventories 201
4.1 Types
to all the sectors of the economy shown in Figure 4–1, but are most applicable to
the manufacturing, wholesale, and retail sectors, which compose approximately
82 percent of the total. The trillion-dollar investment in inventories accounts for
between 20 and 25 percent of the total annual GNP. Clearly there is enormous potential
for improving the efficiency of our economy by intelligently controlling inventories.
Companies that use scientific inventory control methods have a significant competitive
advantage in the marketplace.
A major portion of this text is devoted to presenting and analyzing several mathematical models that can assist with controlling the replenishment of inventories. Both Chapters 4 and 5 assume that the demand for the item is external to the system. In most cases,
this means that the inventory is being acquired or produced to meet the needs of a customer. In a manufacturing environment, however, demands for certain parts are the result
of production schedules for higher-level assemblies; the production-lot-sizing decisions
at one level of the system result in the demand patterns at other levels. The interaction of
components, subassemblies, and final products plays an important role in determining
future demand. Systems of this type are referred to as materials requirements planning
(MRP) systems or dependent demand systems. MRP is treated in detail in Chapter 8.
The fundamental problem of inventory management can be succinctly described by
the two questions (1) When should an order be placed? and (2) How much should be ordered? The complexity of the resulting model depends upon the assumptions one makes
about the various parameters of the system. The major distinction is between models
that assume known demand (this chapter) and those that assume random demand
(Chapter 5), although, as we will see, the form of the cost functions and the assumptions
one makes about physical characteristics of the system also play an important role in
determining the complexity of the resulting model.
In general, the models that we discuss can be used interchangeably to describe either
replenishment from an outside vendor or internal production. This means that from the
point of view of the model, inventory control and production planning are often
synonymous. For example, the lot-sizing methods treated in Chapter 8 could just as well
have been included in this chapter. The issue is not the label that is placed on a technique, but whether it is being correctly applied to the problem being addressed.
4.1
TYPES OF INVENTORIES
When we consider inventories in the context of manufacturing and distribution, there
is a natural classification scheme suggested by the value added from manufacturing or
processing. (This certainly is not the only means of categorizing inventories, but it is
the most natural one for manufacturing applications.)
1. Raw materials. These are the resources required in the production or processing
activity of the firm.
2. Components. Components correspond to items that have not yet reached completion
in the production process. Components are sometimes referred to as subassemblies.
3. Work-in-process. Work-in-process (WIP) is inventory either waiting in the system
for processing or being processed. Work-in-process inventories include component
inventories and may include some raw materials inventories as well. The level of
work-in-process inventory is often used as a measure of the efficiency of a production scheduling system. The just-in-time approach, discussed in detail in Chapter 8,
is aimed at reducing WIP to a minimum.
nah77856_ch04.indd 201
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 202
202
Chapter Four
Inventory Control
202 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
4. Finished goods. Also known as end items, these are the final products of the production process. During production, value is added to the inventory at each level of
the manufacturing operation, culminating with finished goods.
The appropriate label to place on inventory depends upon the context. For example,
components for some operations might be the end products for others.
4.2
MOTIVATION FOR HOLDING INVENTORIES
1. Economies of scale. Consider a company that produces a line of similar items, such
as air filters for automobiles. Each production run of a particular size of filter requires
that the production line be reconfigured and the machines recalibrated. Because the company must invest substantial time and money in setting up to produce each filter size,
enough filters should be produced at each setup to justify this cost. This means that it
could be economical to produce a relatively large number of items in each production run
and store them for future use. This allows the firm to amortize fixed setup costs over a
larger number of units.2
2. Uncertainties. Uncertainty often plays a major role in motivating a firm to store inventories. Uncertainty of external demand is the most important. For example, a retailer
stocks different items so that he or she can be responsive to consumer preferences. If a
customer requests an item that is not available immediately, it is likely that the customer
will go elsewhere. Worse, the customer may never return. Inventory provides a buffer
against the uncertainty of demand.
Other uncertainties provide a motivation for holding inventories as well. One is the uncertainty of the lead time. Lead time is defined as the amount of time that elapses from
the point that an order is placed until it arrives. In the production planning context, interpret the lead time as the time required to produce the item. Even when future demand can
be predicted accurately, the company needs to hold buffer stocks to ensure a smooth flow
of production or continued sales when replenishment lead times are uncertain.
A third significant source of uncertainty is the supply. The OPEC oil embargo of the
late 1970s is an example of the chaos that can result when supply lines are threatened.
Two industries that relied (and continue to rely) heavily on oil and gasoline are the
electric utilities and the airlines. Firms in these and other industries risked having to
curtail operations because of fuel shortages.
Additional uncertainties that could motivate a firm to store inventory include the uncertainty in the supply of labor, the price of resources, and the cost of capital.
3. Speculation. If the value of an item or natural resource is expected to increase, it
may be more economical to purchase large quantities at current prices and store the
items for future use than to pay the higher prices at a future date. In the early 1970s, for
example, the Westinghouse Corporation sustained severe losses on its contracts to
build nuclear plants for several electric utility companies because it guaranteed to supply the uranium necessary to operate the plants at a fixed price. Unfortunately for Westinghouse, the price of the uranium skyrocketed between the time the contracts were
signed and the time the plants were built.
Other industries require large quantities of costly commodities that have experienced considerable fluctuation in price. For example, silver is required for the
production of photographic film. By correctly anticipating a major price increase in
2
This argument assumes that the setup cost is a fixed constant. In some circumstances it can be reduced,
thus justifying smaller lot sizes. This forms the basis of the just-in-time philosophy discussed in detail in
Chapter 8.
nah77856_ch04.indd 202
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 203
4.3 Characteristics of
of Inventory
Inventory Systems 203
Systems 203
4.3 Characteristics
silver, a major producer of photographic film, such as Kodak, could purchase and store
large quantities of silver in advance of the increase and realize substantial savings.
The speculative motive also can be a factor for a firm facing the possibility of a
labor strike. The cost of production could increase significantly when there is a severe
shortage of labor.
4. Transportation. In-transit or pipeline inventories exist because transportation
times are positive. When transportation times are long, as is the case when transporting oil from the Middle East to the United States, the investment in pipeline
inventories can be substantial. One of the disadvantages of producing overseas is the
increased transportation time, and hence the increase in pipeline inventories. This
factor has been instrumental in motivating some firms to establish production operations domestically.
5. Smoothing. Changes in the demand pattern for a product can be deterministic or
random. Seasonality is an example of a deterministic variation, while unanticipated
changes in economic conditions can result in random variation. Producing and storing
inventory in anticipation of peak demand can help to alleviate the disruptions caused
by changing production rates and workforce levels. Smoothing costs and planning for
anticipated swings in the demand were considered in the aggregate planning models in
Chapter 3.
6. Logistics. We use the term logistics to describe reasons for holding inventory different from those already outlined. Certain constraints can arise in the purchasing, production, or distribution of items that force the system to maintain inventory. One such
case is an item that must be purchased in minimum quantities. Another is the logistics
of manufacture; it is virtually impossible to reduce all inventories to zero and expect
any continuity in a manufacturing process.
7. Control costs. An important issue, and one that often is overlooked, is the cost of
maintaining the inventory control system. A system in which more inventory is carried
does not require the same level of control as one in which inventory levels are kept to
a bare minimum. It can be less costly to the firm in the long run to maintain large
inventories of inexpensive items than to expend worker time to keep detailed records
for these items. Even though control costs could be a major factor in determining the
suitability of a particular technique or system, they are rarely factored into the types of
inventory models we will be discussing.
4.3
CHARACTERISTICS OF INVENTORY SYSTEMS
1. Demand. The assumptions one makes about the pattern and characteristics of the
demand often turn out to be the most significant in determining the complexity of the
resulting control model.
a. Constant versus variable. The simplest inventory models assume that the rate of
demand is a constant. The economic order quantity (EOQ) model and its extensions are based on this assumption. Variable demand arises in a variety of contexts,
including aggregate planning (Chapter 3) and materials requirements planning
(Chapter 8).
b. Known versus random. It is possible for demand to be constant in expectation but
still be random. Synonyms for random are uncertain and stochastic. Virtually
all stochastic demand models assume that the average demand rate is constant.
Random demand models are generally both more realistic and more complex
than their deterministic counterparts.
nah77856_ch04.indd 203
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 204
204
Chapter Four
Inventory Control
204 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
2. Lead time. If items are ordered from the outside, the lead time is defined as the
amount of time that elapses from the instant that an order is placed until it arrives. If
items are produced internally, however, then interpret lead time as the amount of time
required to produce a batch of items. We will use the Greek letter to represent lead
time, which is expressed in the same units of time as demand. That is, if demand is
expressed in units per year, then lead time should be expressed in years.
3. Review time. In some systems the current level of inventory is known at all times.
This is an accurate assumption when demand transactions are recorded as they occur.
One example of a system in which inventory levels are known at all times is a modern
supermarket with a visual scanning device at the checkout stand that is linked to a
storewide inventory database. As an item is passed through the scanner, the transaction
is recorded in the database, and the inventory level is decreased by one unit. We will refer
to this case as continuous review. In the other case, referred to as periodic review, inventory levels are known only at discrete points in time. An example of periodic review is a
small grocery store in which physical stock-taking is required to determine the current
levels of on-hand inventory.
4. Excess demand. Another important distinguishing characteristic is how the system
reacts to excess demand (that is, demand that cannot be filled immediately from stock).
The two most common assumptions are that excess demand is either back-ordered (held
over to be satisfied at a future time) or lost (generally satisfied from outside the system).
Other possibilities include partial back-ordering (part of the demand is back-ordered and
part of the demand is lost) or customer impatience (if the customer’s order is not filled
within a fixed amount of time, he or she cancels). The vast majority of inventory models,
especially the ones that are used in practice, assume full back-ordering of excess demand.
5. Changing inventory. In some cases the inventory undergoes changes over time that
may affect its utility. Some items have a limited shelf life, such as food, and others may
become obsolete, such as automotive spare parts. Mathematical models that
incorporate the effects of perishability or obsolescence are generally quite complex and
beyond the scope of this text. A brief discussion can be found in Section 5.8.
4.4
RELEVANT COSTS
Because we are interested in optimizing the inventory system, we must determine an
appropriate optimization or performance criterion. Virtually all inventory models use
cost minimization as the optimization criterion. An alternative performance criterion
might be profit maximization. However, cost minimization and profit maximization are
essentially equivalent criteria for most inventory control problems. Although different
systems have different characteristics, virtually all inventory costs can be placed into
one of three categories: holding cost, order cost, or penalty cost. We discuss each in turn.
Holding Cost
The holding cost, also known as the carrying cost or the inventory cost, is the sum of all
costs that are proportional to the amount of inventory physically on hand at any point
in time. The components of the holding cost include a variety of seemingly unrelated
items. Some of these are
•
•
•
•
nah77856_ch04.indd 204
Cost of providing the physical space to store the items.
Taxes and insurance.
Breakage, spoilage, deterioration, and obsolescence.
Opportunity cost of alternative investment.
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 205
4.4 Relevant Costs 205
Costs 205
4.4 Relevant
The last item often turns out to be the most significant in computing holding costs
for most applications. Inventory and cash are in some sense equivalent. Capital must
be invested to either purchase or produce inventory, and decreasing inventory levels
results in increased capital. This capital could be invested by the company either
internally, in its own operation, or externally.
What is the interest rate that could be earned on this capital? You and I can place our
money in a simple passbook account with an interest rate of 2 percent, or possibly a
long-term certificate of deposit with a return of maybe 5 percent. We could earn somewhat more by investing in high-yield bond funds or buying short-term industrial paper
or second deeds of trust.
In general, however, most companies must earn higher rates of return on their
investments than do individuals in order to remain profitable. The value of the interest rate that corresponds to the opportunity cost of alternative investment is related
to (but not the same as) a number of standard accounting measures, including the internal rate of return, the return on assets, and the hurdle rate (the minimum rate that
would make an investment attractive to the firm). Finding the right interest rate for
the opportunity cost of alternative investment is very difficult. Its value is estimated
by the firm’s accounting department and is usually an amalgam of the accounting
measures listed earlier. For convenience, we will use the term cost of capital to refer
to this component of the holding cost. We may think of the holding cost as an aggregated interest rate comprised of the four components we listed. For example,
28% � Cost of capital
2% � Taxes and insurance
6% � Cost of storage
1% � Breakage and spoilage
37% � Total interest charge
This would be interpreted as follows: We would assess a charge of 37 cents for every
dollar that we have invested in inventory during a one-year period. However, as we generally measure inventory in units rather than in dollars, it is convenient to express the
holding cost in terms of dollars per unit per year rather than dollars per dollar per year.
Let c be the dollar value of one unit of inventory, I be the annual interest rate, and h be
the holding cost in terms of dollars per unit per year. Then we have the relationship
h � Ic.
Hence, in this example, an item valued at $180 would have an annual holding cost
of h � (0.37)($180) � $66.60. If we held 300 of these items for five years, the total
holding cost over the five years would be
(5)(300)(66.60) � $99,900.
This example raises an interesting question. Suppose that during the five-year
period the inventory level did not stay fixed at 300 but varied on a continuous basis. We
would expect inventory levels to change over time. Inventory levels decrease when
items are used to satisfy demand and increase when units are produced or new orders
arrive. How would the holding cost be computed in such a case? In particular, suppose
the inventory level I(t) during some interval (t1, t2) behaves as in Figure 4–2.
The holding cost incurred at any point in time is proportional to the inventory level
at that point in time. In general, the total holding cost incurred from a time t1 to a time
t2 is h multiplied by the area under the curve described by I(t). The average inventory
nah77856_ch04.indd 205
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 206
206
Chapter Four
Inventory Control
206 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
FIGURE 4–2
I(t)
Inventory as a function
of time
Average
inventory
level
during t1 to t2
t1
t2
t
level during the period (t1, t2) is the area under the curve divided by t2 � t1. For the
cases considered in this chapter, simple geometry can be used to find the area under the
inventory level curve. When I(t) is described by a straight line, its average value is
obvious. In cases such as that pictured in Figure 4–2, in which the curve of I(t) is complex, the average inventory level would be determined by computing the integral of I(t)
over the interval (t1, t2) and dividing by t2 � t1.
Order Cost
The holding cost includes all those costs that are proportional to the amount of inventory on hand, whereas the order cost depends on the amount of inventory that is ordered
or produced.
In most applications, the order cost has two components: a fixed and a variable component. The fixed cost, K, is incurred independent of the size of the order as long as it is
not zero. The variable cost, c, is incurred on a per-unit basis. We also refer to K as the
setup cost and c as the proportional order cost. Define C(x) as the cost of ordering (or
producing) x units. It follows that
C(x) � b
0
K � cx
if x � 0,
if x 7 0.
The order cost function is pictured in Figure 4–3.
When estimating the setup cost, one should include only those costs that are relevant
to the current ordering decision. For example, the cost of maintaining the purchasing
department of the company is not relevant to daily ordering decisions and should not be
factored into the estimation of the setup cost. This is an overhead cost that is independent of the decision of whether or not an order should be placed. The appropriate costs
comprising K would be the bookkeeping expense associated with the order, the fixed
costs independent of the size of the order that might be required by the vendor, costs of
order generation and receiving, and handling costs.
nah77856_ch04.indd 206
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 207
4.4 Relevant Costs 207
Costs 207
4.4 Relevant
FIGURE 4–3
Cost of ordering x units
Order cost function
Slope = c
K
x
Penalty Cost
The penalty cost, also known as the shortage cost or the stock-out cost, is the cost
of not having sufficient stock on hand to satisfy a demand when it occurs. This cost has
a different interpretation depending on whether excess demand is back-ordered (orders that cannot be filled immediately are held on the books until the next shipment
arrives) or lost (known as lost sales). In the back-order case, the penalty cost includes
whatever bookkeeping and/or delay costs might be involved. In the lost-sales case, it
includes the lost profit that would have been made from the sale. In either case, it would
also include the loss-of-goodwill cost, which is a measure of customer satisfaction.
Estimating the loss-of-goodwill component of the penalty cost can be very difficult in
practice.
We use the symbol p to denote penalty cost and assume that p is charged on a perunit basis. That is, each time a demand occurs that cannot be satisfied immediately, a
cost p is incurred independent of how long it takes to eventually fill the demand. An
alternative means of accounting for shortages is to charge the penalty cost on a
per-unit-per-unit-time basis (as we did with the holding cost). This approach is appropriate if the time that a back order stays on the books is important, for example, if a
back order results in stopping a production line because of the unavailability of a
part. The models considered in this chapter assume that penalty costs are charged on
a per-unit basis only. Penalty cost models are not considered in this chapter. Penalty
costs are included in Chapter 5, but models incorporating a time-weighted penalty cost
are not.
We present those inventory models that have had the greatest impact in the user
community. Many of the techniques discussed in both this chapter and in Chapter 5
form the basis for commercial inventory control systems or in-house systems. In most
cases, the models are simple enough that optimal operating policies can be calculated
by hand, but they are often complex enough to capture the essential trade-offs in
inventory management.
nah77856_ch04.indd 207
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 208
208
Chapter Four
Inventory Control
208 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
Problems for Sections 4.1–4.4
1. What are the two questions that inventory control addresses?
2. Discuss the cost penalties incurred by a firm that holds too much inventory and
one that holds too little inventory.
3. ABC, Inc., produces a line of touring bicycles. Specifically, what are the four types
of inventories (raw materials, components, work-in-process, and finished goods)
that would arise in the production of this item?
4. I Carry rents trucks for moving and hauling. Each truck costs the company an
average of $8,000, and the inventory of trucks varies monthly depending on the
number that are rented out. During the first eight months of last year, I Carry had
the following ending inventory of trucks on hand:
Month
January
February
March
April
Number of Trucks
26
38
31
22
Month
May
June
July
August
Number of Trucks
13
9
16
5
I Carry uses a 20 percent annual interest rate to represent the cost of capital.
Yearly costs of storage amount to 3 percent of the value of each truck, and the cost
of liability insurance is 2 percent.
a. Determine the total handling cost incurred by I Carry during the period January
to August. Assume for the purposes of your calculation that the holding cost
incurred in a month is proportional to the inventory on hand at the end of the
month.
b. Assuming that these eight months are representative, estimate the average annual cost of holding trucks.
5. Stationery Supplies is considering installing an inventory control system in its
store in Provo, Utah. The store carries about 1,400 different inventory items and
has annual gross sales of about $80,000. The inventory control system would cost
$12,500 to install and about $2,000 per year in additional supplies, time, and
maintenance. If the savings to the store from the system can be represented as a
fixed percentage of annual sales, what would that percentage have to be in order
for the system to pay for itself in five years or less?
6. For Stationery Supplies, discussed in Problem 5, list and discuss all the uncertainties
that would motivate the store to maintain inventories of its 1,400 items.
7. Stationery Supplies orders plastic erasers from a company in Nürnberg, Germany.
It takes six weeks to ship the erasers from Germany to Utah. Stationery Supplies
maintains a standing order of 200 erasers every six months (shipped on the first of
January and the first of July).
a. Assuming the ordering policy the store is using does not result in large buildups
of inventory or long-term stock-outs, what is the annual demand for erasers?
b. Draw a graph of the pipeline inventory (that is, the inventory ordered but not
received) of the erasers during one year. What is the average pipeline inventory
of erasers during the year?
nah77856_ch04.indd 208
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 209
4.4 Relevant Costs 209
Costs 209
4.4 Relevant
c. Express the replenishment lead time in years and multiply the annual demand
you obtained in part (a) by the lead time. What do you notice about the result
that you obtain?
8. Penalty costs can be assessed only against the number of units of demand that cannot be satisfied, or against the number of units weighted by the amount of time that
an order stays on the books. Consider the following history of supply and demand
transactions for a particular part:
Month
Number of Items
Received
January
February
March
April
May
June
200
175
750
950
500
2,050
Demand during
Month
520
1,640
670
425
280
550
Assume that starting inventory at the beginning of January is 480 units.
a. Determine the ending inventory each month. Assume that excess demands are
back-ordered.
b. Assume that each time a unit is demanded that cannot be supplied immediately,
a one-time charge of $10 is made. Determine the stock-out cost incurred during
the six months (1) if excess demand at the end of each month is lost, and (2) if
excess demand at the end of each month is back-ordered.
c. Suppose that each stock-out costs $10 per unit per month that the demand
remains unfilled. If demands are filled on a first-come, first-served basis, what
is the total stock-out cost incurred during the six months using this type of cost
criterion? (Assume that the demand occurs at the beginning of the month for
purposes of your calculation.) Notice that you must assume that excess
demands are back-ordered for this case to make any sense.
d. Discuss under what circumstances the cost criterion used in part (b) might be
appropriate and under what circumstances the cost criterion used in part (c)
might be appropriate.
9. HAL Ltd. produces a line of high-capacity disk drives for mainframe computers.
The housings for the drives are produced in Hamilton, Ontario, and shipped to
the main plant in Toronto. HAL uses the drive housings at a fairly steady rate of 720
per year. Suppose that the housings are shipped in trucks that can hold 40 housings
at one time. It is estimated that the fixed cost of loading the housings onto the truck
and unloading them on the other end is $300 for shipments of 120 or fewer housings
(i.e., three or fewer truckloads). Each trip made by a single truck costs the company
$160 in driver time, gasoline, oil, insurance, and wear and tear on the truck.
a. Compute the annual costs of transportation and loading and unloading
the housings for the following policies: (1) shipping one truck per week,
(2) shipping one full truckload as often as needed, and (3) shipping three full
truckloads as often as needed.
b. For what reasons might the policy in part (a) with the highest annual cost be
more desirable from a systems point of view than the policy having the lowest
annual cost?
nah77856_ch04.indd 209
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 210
210
Chapter Four
Inventory Control
210 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
4.5
THE EOQ MODEL
The EOQ model (or economic order quantity model) is the simplest and most fundamental of all inventory models. It describes the important trade-off between fixed
order costs and holding costs, and is the basis for the analysis of more complex
systems.
The Basic Model
Assumptions:
1. The demand rate is known and is a constant units per unit time. (The unit of time
may be days, weeks, months, etc. In what follows we assume that the default unit
of time is a year. However, the analysis is valid for other time units as long as all
relevant variables are expressed in the same units.)
2. Shortages are not permitted.
3. There is no order lead time. (This assumption will be relaxed.)
4. The costs include
a. Setup cost at K per positive order placed.
b. Proportional order cost at c per unit ordered.
c. Holding cost at h per unit held per unit time.
Assume with no loss in generality that the on-hand inventory at time zero is zero.
Shortages are not allowed, so we must place an order at time zero. Let Q be the size of
the order. It follows that the on-hand inventory level increases instantaneously from
zero to Q at time t � 0.
Consider the next time an order is to be placed. At this time, either the inventory is positive or it is again zero. A little reflection shows that we can reduce the holding costs by
waiting until the inventory level drops to zero before ordering again.At the instant that onhand inventory equals zero, the situation looks exactly the same as it did at time t � 0. If
it was optimal to place an order for Q units at that time, then it is still optimal to order Q
units. It follows that the function that describes the changes in stock levels over time is the
familiar sawtooth pattern of Figure 4–4.
The objective is to choose Q to minimize the average cost per unit time. Unless
otherwise stated, we will assume that a unit of time is a year, so that we minimize
the average annual cost. Other units of time, such as days, weeks, or months, are also
acceptable, as long as all time-related variables are expressed in the same units. One
might think that the appropriate optimization criterion would be to minimize the total
cost in a cycle. However, this ignores the fact that the cycle length itself is a function
of Q and must be explicitly included in the formulation.
Next, we derive an expression for the average annual cost as a function of the lot
size Q. In each cycle, the total fixed plus proportional order cost is C(Q) � K � cQ. In
order to obtain the order cost per unit time, we divide by the cycle length T. As Q units
are consumed each cycle at a rate , it follows that T � Q. This result also can be
obtained by noting that the slope of the inventory curve pictured in Figure 4–4, �,
equals the ratio �QT.
Consider the holding cost. Because the inventory level decreases linearly from
Q to 0 each cycle, the average inventory level during one order cycle is Q2. Because
all cycles are identical, the average inventory level over a time horizon composed
nah77856_ch04.indd 210
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 211
4.5 The EOQ
EOQ Model 211
Model 211
4.5 The
FIGURE 4–4
Slope = – Q
Inventory (I(t))
Inventory levels for
the EOQ model
Time t
T
of many cycles is also Q2. It follows that the average annual cost, say G(Q), is
given by
K � cQ hQ K � cQ hQ
�
�
�
T
2
Q>
2
hQ
K
�
� c �
.
Q
2
G(Q) �
The three terms composing G(Q) are annual setup cost, annual purchase cost, and
annual holding cost, respectively.
We now wish to find Q to minimize G(Q). Consider the shape of the curve G(Q). We
have that
G(Q) � �KQ2 � h2
and
G(Q) � 2KQ3 � 0
for Q � 0.
Since G(Q) � 0, it follows that G(Q) is a convex function of Q. Furthermore, since
G(0) � �� and G(�) � h2, it follows that G(Q) behaves as pictured in Figure 4–5.
The optimal value of Q occurs where G(Q) � 0. This is true when Q2 � 2Kh,
which gives
2K
Q* �
.
A h
Q* is known as the economic order quantity (EOQ). There are a number of interesting
points to note:
1. In Figure 4–5, the curves corresponding to the fixed order cost component KQ
and the holding cost component hQ2 also are included. Notice that Q* is the value of
nah77856_ch04.indd 211
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 212
212
Chapter Four
Inventory Control
212 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
FIGURE 4–5
The average annual
cost function G(Q)
G(Q )
600
500
400
300
G(Q)
hQ
2
200
100
0
1
5
10
15
20
K�
Q
25
30
Q
Q
Q where the two curves intersect. (If you equate hQ2 and KQ and solve for Q, you
will obtain the EOQ formula.) In general, the minimum of the sum of two functions
will not occur at the intersection of the two functions. It is an interesting coincidence
that it does in this case.
2. Notice that the proportional order cost component, c, does not appear explicitly in
the expression for Q*. This is because the term c appearing in the definition of G(Q) is
independent of Q. As all feasible policies must replenish inventory at the rate of demand,
the proportional order cost incurred per unit time is c independent of Q. Because c is
a constant, we generally ignore it when computing average costs. Notice that c does
affect the value of Q* indirectly, as h appears in the EOQ formula and h � Ic.
Example 4.1
Number 2 pencils at the campus bookstore are sold at a fairly steady rate of 60 per week. The
pencils cost the bookstore 2 cents each and sell for 15 cents each. It costs the bookstore $12
to initiate an order, and holding costs are based on an annual interest rate of 25 percent.
Determine the optimal number of pencils for the bookstore to purchase and the time between
placement of orders. What are the yearly holding and setup costs for this item?
Solution
First, we convert the demand to a yearly rate so that it is consistent with the interest charge,
which is given on an annual basis. (Alternatively, we could have converted the annual interest
rate to a weekly interest rate.) The annual demand rate is � (60)(52) � 3,120. The holding
cost h is the product of the annual interest rate and the variable cost of the item. Hence, h �
(0.25)(0.02) � 0.005. Substituting into the EOQ formula, we obtain
Q* �
2K
B h
�
B
(2)(12)(3,120)
� 3,870.
0.005
The cycle time is T � Q � 3,8703,120 � 1.24 years. The average annual holding cost is
h(Q2) � 0.005(3,8702) � $9.675. The average annual setup cost is KQ, which is also
$9.675.
nah77856_ch04.indd 212
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 213
4.5 The EOQ
EOQ Model 213
Model 213
4.5 The
Example 4.1 illustrates some of the problems that can arise when using simple models. The optimal solution calls for ordering almost 4,000 pencils every 15 months.
Even though this value of Q minimizes the yearly holding and setup costs, it could be
infeasible: the store may not have the space to store 4,000 pencils. Simple models cannot account for all the constraints present in a real problem. For that reason, every solution must be considered in context and modified, if necessary, to fit the application.
Notice also that the optimal solution did not depend on the selling price of 15 cents.
Even if each pencil sold for $2, we would recommend the same order quantity, because
the pencils are assumed to sell at a rate of 60 per week no matter what their price. This is,
of course, a simplification of reality. It is reasonable to assume that the demand is relatively stable for a range of prices. Inventory models explicitly incorporate selling price in
the formulation only when the selling price is included as part of the optimization.
Inclusion of Order Lead Time
One of the assumptions made in our derivation of the EOQ model was that there was
no order lead time. We now relax that assumption. Suppose in Example 4.1 that the
pencils had to be ordered four months in advance. If we were to place the order exactly
four months before the end of the cycle, the order would arrive at exactly the same
point in time as in the zero lead time case. The optimal timing of order placement for
Example 4.1 is shown in Figure 4–6.
Rather than say that an order should be placed so far in advance of the end of a
cycle, it is more convenient to indicate reordering in terms of the on-hand inventory.
Define R, the reorder point, as the level of on-hand inventory at the instant an order
should be placed. From Figure 4–6, we see that R is the product of the lead time and
the demand rate (R ). For the example, R (3,120)(0.3333) 1,040. Notice that
we converted the lead time to years before multiplying. Always express all relevant
variables in the same units of time.
Determining the reorder point is more difficult when the lead time exceeds a cycle.
Consider an item with an EOQ of 25, a demand rate of 500 units per year, and a lead
time of six weeks. The cycle time is T 25500 0.05 year, or 2.6 weeks. Forming
FIGURE 4–6
I(t)
Reorder point
calculation for
Example 4.1
Q = 3,870
1.24 years
4
months
R = 1,040
Order placed
t
Order arrives
nah77856_ch04.indd 213
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 214
214
Chapter Four
Inventory Control
214 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
Reorder point
calculation for lead
times exceeding one
cycle
I(t)
FIGURE 4–7
Q = 25
.31
cycle =
.0155
year
R=8
Order placed
t
Order arrives
2.31 cycles = .1154 year
the ratio of T, we obtain 2.31. This means that there are exactly 2.31 cycles in the
lead time. Every order must be placed 2.31 cycles in advance (see Figure 4–7).
Notice that for the purpose of computing the reorder point, this is exactly the same
as placing the order 0.31 cycle in advance. This is true because the level of on-hand
inventory is the same whether we are at a point 2.31 or 0.31 cycle before the arrival of
an order. In this case, 0.31 cycle is 0.0155 year, thus giving a reorder point of R (0.0155)(500) 7.75 8. In general, when T, use the following procedure:
a. Form the ratio T.
b. Consider only the fractional remainder of the ratio. Multiply this fractional
remainder by the cycle length to convert back to years.
c. Multiply the result of step (b) by the demand rate to obtain the reorder point.
Sensitivity
In this part we examine the issue of how sensitive the annual cost function is to errors
in the calculation of Q. Consider Example 4.1. Suppose that the bookstore orders
pencils in batches of 1,000, rather than 3,870 as the optimal solution indicates. What
additional cost is it incurring by using a suboptimal solution? To answer the question,
we consider the average annual cost function G(Q). By substituting Q 1,000, we can
find the average annual cost for this lot size and compare it to the optimal cost to
determine the magnitude of the penalty. We have
G(Q) KQ hQ2 (12)(3,120)1,000 (0.005)(1,000)2 $39.94,
which is considerably larger than the optimal cost of $19.35.
One can find the cost penalty for suboptimal solutions in this manner for any
particular problem. However, it is more instructive and more convenient to obtain a
universal solution to the sensitivity problem. We do so by deriving an expression for
the ratio of the suboptimal cost over the optimal cost as a function of the ratio of the
optimal and suboptimal order quantities. Let G* be the average annual holding and
nah77856_ch04.indd 214
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 215
4.5 The EOQ
EOQ Model 215
Model 215
4.5 The
setup cost at the optimal solution. Then
G* K>Q* hQ*>2 K
22K>h
2Kh.
h 2K
Kh
2
2A h
A 2
It follows that for any Q,
G(Q) K>Q hQ>2
G*
22Kh
1
2K Q
h
2Q A h
2 A 2K
Q*
Q
2Q 2Q*
Q
1 Q*
d.
c
2 Q
Q*
To see how one would use this result, consider using a suboptimal lot size in Example 4.1. The optimal solution was Q* 3,870, and we wished to evaluate the cost error
of using Q 1,000. Forming the ratio Q*Q gives 3.87. Hence, G(Q)G* (0.5)(3.87 13.87) (0.5)(4.128) 2.06. This says that the average annual holding
and setup cost with Q 1,000 is 2.06 times the optimal average annual holding and
setup cost.
In general, the cost function G(Q) is relatively insensitive to errors in Q. For example,
if Q is twice as large as it should be, QQ* is 2 and GG* is 1.25. Hence, an error of
100 percent in the value of Q results in an error of only 25 percent in the annual holding
and setup cost. Notice that you obtain the same result if Q is half Q*, since Q*Q 2.
However, this does not imply that the average annual cost function is symmetric. In fact,
suppose that the order quantity differed from the optimal by Q units. A value of Q Q* Q would result in a lower average annual cost than a value of Q Q* Q.
EOQ and JIT
Largely as a result of the success of Toyota’s kanban system, a new philosophy is emerging about the role and importance of inventories in manufacturing environments. This
philosophy, known as just-in-time (JIT), says that excess work-in-process inventories
are not desirable, and inventories should be reduced to the bare essentials. (We discuss
the just-in-time philosophy in more detail in Chapter 1 and the mechanics of kanban in
Chapter 8.) EOQ is the result of traditional thinking about inventories and scale
economies in economics. Are the EOQ and JIT approaches at odds with each other?
Proponents argue that an essential part of implementing JIT is reducing setup
times, and hence setup costs. As setup costs decrease, traditional EOQ theory says that
lot sizes should be reduced. In this sense, the two ways of thinking are compatible.
However, there are times when they may not be. We believe that there is substantial
value to the JIT approach that may not be incorporated easily into a mathematical
model. Quality problems can be identified and rectified before inventories of defective
parts accumulate. Plants can be more flexible if they are not burdened with excess inprocess inventories. Certainly Toyota’s success with the JIT approach, as evidenced by
substantially lower inventory costs per car than are typical for U.S. auto manufacturers, is a testament to the value of JIT.
nah77856_ch04.indd 215
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 216
216
Chapter Four
Inventory Control
216 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
However, we believe that every new approach must be incorporated carefully into
the firm’s business and not adopted blindly without evaluating its consequences and
appropriateness. JIT, although an important development in material management, is
not always the best approach. The principles underlying EOQ (and MRP, discussed
in Chapter 8) are sound and should not be ignored. The following example illustrates
this point.
Example 4.2
The Rahway, New Jersey, plant of Metalcase, a manufacturer of office furniture, produces metal
desks at a rate of 200 per month. Each desk requires 40 Phillips head metal screws purchased
from a supplier in North Carolina. The screws cost 3 cents each. Fixed delivery charges and costs
of receiving and storing shipments of the screws amount to about $100 per shipment, independently of the size of the shipment. The firm uses a 25 percent interest rate to determine
holding costs. Metalcase would like to establish a standing order with the supplier and is considering several alternatives. What standing order size should they use?
Solution
First we compute the EOQ. The annual demand for screws is
(200)(12)(40) 96,000.
The annual holding cost per screw is (0.25)(0.03) 0.0075. From the EOQ formula, the optimal lot size is
Q* B
(2)(100)(96,000)
50,597.
0.0075
Note that the cycle time is T Q 50,59796,000 0.53 year or about once every six
months. Hence the optimal policy calls for replenishment of the screws about twice a year. A
JIT approach would be to order the screws as frequently as possible to minimize the inventory
held at the plant. Implementing such an approach might suggest a policy of weekly deliveries.
Such a policy makes little sense in this context, however. This policy would require 52 deliveries
per year, incurring setup costs of $5,200 annually. The EOQ solution gives a total annual cost
of both setups and holding of less than $400. For a low-value item such as this with high fixed
order costs, small lot sizes in accordance with JIT are inappropriate. The point is that no single
approach should be blindly adopted for all situations. The success of a method in one context
does not ensure its appropriateness in all other contexts.
Problems for Section 4.5
10. A specialty coffeehouse sells Colombian coffee at a fairly steady rate of 280 pounds
annually. The beans are purchased from a local supplier for $2.40 per pound. The coffeehouse estimates that it costs $45 in paperwork and labor to place an order for the
coffee, and holding costs are based on a 20 percent annual interest rate.
a. Determine the optimal order quantity for Colombian coffee.
b. What is the time between placement of orders?
c. What is the average annual cost of holding and setup due to this item?
d. If replenishment lead time is three weeks, determine the reorder level based on
the on-hand inventory.
11. For the situation described in Problem 10, draw a graph of the amount of inventory
on order. Using your graph, determine the average amount of inventory on order.
Also compute the demand during the replenishment lead time. How do these two
quantities differ?
nah77856_ch04.indd 216
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 217
4.5 The EOQ
EOQ Model 217
Model 217
4.5 The
12. A large automobile repair shop installs about 1,250 mufflers per year, 18 percent
of which are for imported cars. All the imported-car mufflers are purchased from
a single local supplier at a cost of $18.50 each. The shop uses a holding cost based
on a 25 percent annual interest rate. The setup cost for placing an order is estimated to be $28.
a. Determine the optimal number of imported-car mufflers the shop should purchase each time an order is placed, and the time between placement of orders.
b. If the replenishment lead time is six weeks, what is the reorder point based on
the level of on-hand inventory?
c. The current reorder policy is to buy imported-car mufflers only once a year.
What are the additional holding and setup costs incurred by this policy?
13. Consider the coffeehouse discussed in Problem 10. Suppose that its setup cost for
ordering was really only $15. Determine the error made in calculating the annual
cost of holding and setup incurred as a result of its using the wrong value of K.
(Note that this implies that its current order policy is suboptimal.)
14. A local machine shop buys hex nuts and molly screws from the same supplier. The
hex nuts cost 15 cents each and the molly screws cost 38 cents each. A setup cost
of $100 is assumed for all orders. This includes the cost of tracking and receiving
the orders. Holding costs are based on a 25 percent annual interest rate. The shop
uses an average of 20,000 hex nuts and 14,000 molly screws annually.
a. Determine the optimal size of the orders of hex nuts and molly screws, and the
optimal time between placement of orders of these two items.
b. If both items are ordered and received simultaneously, the setup cost of $100
applies to the combined order. Compare the average annual cost of holding and
setup if these items are ordered separately; if they are both ordered when the hex
nuts would normally be ordered; and if they are both ordered when the molly
screws would normally be ordered.
15. David’s Delicatessen flies in Hebrew National salamis regularly to satisfy a growing demand for the salamis in Silicon Valley. The owner, David Gold, estimates
that the demand for the salamis is pretty steady at 175 per month. The salamis cost
Gold $1.85 each. The fixed cost of calling his brother in New York and having the
salamis flown in is $200. It takes three weeks to receive an order. Gold’s accountant, Irving Wu, recommends an annual cost of capital of 22 percent, a cost of shelf
space of 3 percent of the value of the item, and a cost of 2 percent of the value for
taxes and insurance.
a. How many salamis should Gold have flown in and how often should he order
them?
b. How many salamis should Gold have on hand when he phones his brother to
send another shipment?
c. Suppose that the salamis sell for $3 each. Are these salamis a profitable item
for Gold? If so, what annual profit can he expect to realize from this item?
(Assume that he operates the system optimally.)
d. If the salamis have a shelf life of only 4 weeks, what is the trouble with the policy that you derived in part (a)? What policy would Gold have to use in that
case? Is the item still profitable?
16. In view of the results derived in the section on sensitivity analysis, discuss the following quotation of an inventory control manager: “If my lot sizes are going to be
off the mark, I’d rather miss on the high side than on the low side.”
nah77856_ch04.indd 217
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 218
218
Chapter Four
Inventory Control
218 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
4.6
EXTENSION TO A FINITE PRODUCTION RATE
An implicit assumption of the simple EOQ model is that the items are obtained from
an outside supplier. When that is the case, it is reasonable to assume that the entire lot
is delivered at the same time. However, if we wish to use the EOQ formula when the
units are produced internally, then we are effectively assuming that the production rate
is infinite. When the production rate is much larger than the demand rate, this
assumption is probably satisfactory as an approximation. However, if the rate of
production is comparable to the rate of demand, the simple EOQ formula will lead to
incorrect results.
Assume that items are produced at a rate P during a production run. We require that
P � for feasibility. All other assumptions will be identical to those made in the derivation of the simple EOQ. When units are produced internally, the curve describing
inventory levels as a function of time is slightly different from the sawtooth pattern of
Figure 4 –4. The change in the inventory level over time for the finite production rate
case is shown in Figure 4–8.
Let Q be the size of each production run. Let T, the cycle length, be the time
between successive production startups. Write T � T1 � T2, where T1 is uptime (production time) and T2 is downtime. Note that the maximum level of on-hand inventory
during a cycle is not Q.
The number of units consumed each cycle is T, which must be the same as the
number of units produced each cycle, which is simply Q. It follows that Q � T, or
T � Q. Define H as the maximum level of on-hand inventory. As items are produced
at a rate P for a time T1, it follows that Q � PT1, or T1 � QP. From Figure 4–8 we see
that HT1 � P � . This follows from the definition of the slope as the rise over the
run. Substituting T1 � QP and solving for H gives H � Q(1 � P).
We now determine an expression for the average annual cost function. Because the
average inventory level is H2, it follows that
G(Q) �
K hH K hQ
�
�
�
(1 � l>P).
T
2
Q
2
Inventory levels for
finite production rate
model
I(t)
FIGURE 4–8
Slope = P – H
Slope = – T1
nah77856_ch04.indd 218
T
T2
t
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 219
4.6 Extension to
to aa Finite
Finite Production
Production Rate 219
Rate 219
4.6 Extension
Notice that if we define h� � h(1 � �P), then this G(Q) is identical to that of the
infinite production rate case with h� substituted for h. It follows that
Q* �
2K�
.
A h¿
Example 4.3
A local company produces an erasable programmable read-only memory (EPROM) for several
industrial clients. It has experienced a relatively flat demand of 2,500 units per year for the
product. The EPROM is produced at a rate of 10,000 units per year. The accounting department
has estimated that it costs $50 to initiate a production run, each unit costs the company $2 to
manufacture, and the cost of holding is based on a 30 percent annual interest rate. Determine
the optimal size of a production run, the length of each production run, and the average annual
cost of holding and setup. What is the maximum level of the on-hand inventory of the EPROMs?
Solution
First, we compute h � (0.3)(2) � 0.6 per unit per year. The modified holding cost is h� �
h(1 � �P) � (0.6)(1 � 2,50010,000) � 0.45. Substituting into the EOQ formula and using h�
for h, we obtain Q* � 745. Note that the simple EOQ equals 645, all out 14 percent less.
The time between production runs is T � Q� � 7452,500 � 0.298 year. The uptime each
cycle is T1 � QP � 74510,000 � 0.0745 year, and the downtime each cycle is T2 � T � T1 �
0.2235 year.
The average annual cost of holding and setup is
G(Q*) �
h�Q*
(50)(2,500)
(0.45)(745)
K�
�
�
�
� 335.41.
Q*
2
745
2
The maximum level of on-hand inventory is H � Q*(1 � �P) � 559 units.
Problems for Section 4.6
17. The Wod Chemical Company produces a chemical compound that is used as a lawn
fertilizer. The compound can be produced at a rate of 10,000 pounds per day. Annual
demand for the compound is 0.6 million pounds per year. The fixed cost of setting
up for a production run of the chemical is $1,500, and the variable cost of production is $3.50 per pound. The company uses an interest rate of 22 percent to account
for the cost of capital, and the costs of storage and handling of the chemical amount
to 12 percent of the value. Assume that there are 250 working days in a year.
a. What is the optimal size of the production run for this particular compound?
b. What proportion of each production cycle consists of uptime and what proportion
consists of downtime?
c. What is the average annual cost of holding and setup attributed to this item? If
the compound sells for $3.90 per pound, what is the annual profit the company
is realizing from this item?
18. Determine the batch size that would result in Problem 17 if you assumed that the
production rate was infinite. What is the additional average annual cost that would
be incurred using this batch size rather than the one you found in Problem 17?
19. HAL Ltd., discussed in Problem 9, can produce the disk drive housings in the
Hamilton, Ontario, plant at a rate of 150 housings per month. The housings cost
HAL $85 each to produce, and the setup cost for beginning a production run is $700.
Assume an annual interest rate of 28 percent for determining the holding cost.
a. What is the optimal number of housings for HAL to produce in each production run?
nah77856_ch04.indd 219
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 220
220
Chapter Four
Inventory Control
220 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
b. Find the time between initiation of production runs, the time devoted to production, and the downtime each production cycle.
c. What is the maximum dollar investment in housings that HAL has at any point
in time?
20. Filter Systems produces air filters for domestic and foreign cars. One filter, part
number JJ39877, is supplied on an exclusive contract basis to Oil Changers at a
constant 200 units monthly. Filter Systems can produce this filter at a rate of 50 per
hour. Setup time to change the settings on the equipment is 1.5 hours. Worker time
(including overhead) is charged at the rate of $55 per hour, and plant idle time
during setups is estimated to cost the firm $100 per hour in lost profit.
Filter Systems has established a 22 percent annual interest charge for determining
holding cost. Each filter costs the company $2.50 to produce; they are sold for
$5.50 each to Oil Changers. Assume 6-hour days, 20 working days per month, and
12 months per year for your calculations.
a. How many JJ39877 filters should Filter Systems produce in each production
run of this particular part to minimize annual holding and setup costs?
b. Assuming that it produces the optimal number of filters in each run, what is the
maximum level of on-hand inventory of these filters that the firm has at any
point in time?
c. What percentage of the working time does the company produce these particular filters, assuming that the policy in part (a) is used?
4.7
QUANTITY DISCOUNT MODELS
We have assumed up until this point that the cost c of each unit is independent of the
size of the order. Often, however, the supplier is willing to charge less per unit for larger
orders. The purpose of the discount is to encourage the customer to buy the product in
larger batches. Such quantity discounts are common for many consumer goods.
Although many different types of discount schedules exist, there are two that seem
to be the most popular: all-units and incremental. In each case we assume that there are
one or more breakpoints defining changes in the unit cost. However, there are two
possibilities: either the discount is applied to all the units in an order (all-units), or it is
applied only to the additional units beyond the breakpoint (incremental). The all-units
case is more common.
Example 4.4
The Weighty Trash Bag Company has the following price schedule for its large trash can liners.
For orders of less than 500 bags, the company charges 30 cents per bag; for orders of 500 or
more but fewer than 1,000 bags, it charges 29 cents per bag; and for orders of 1,000 or more,
it charges 28 cents per bag. In this case the breakpoints occur at 500 and 1,000. The discount
schedule is all-units because the discount is applied to all of the units in an order. The order cost
function C(Q) is defined as
0.30Q
C(Q) 0.29Q
0.28Q
for
for
for
0 Q 500,
500 Q 1,000,
1,000 Q
The function C(Q) is pictured in Figure 4–9. In Figure 4–10, we consider the same breakpoints, but assume an incremental quantity discount schedule.
Note that the average cost per unit with an all-units schedule will be less than the
average cost per unit with the corresponding incremental schedule.
nah77856_ch04.indd 220
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 221
4.7 Quantity Discount
Discount Models 221
Models 221
4.7 Quantity
FIGURE 4–9
C(Q)
All-units discount
order cost function
c0 = .30
500
c1 = .29
c2 = .28
Q
1,000
FIGURE 4–10
Incremental discount
order cost function
C(Q)
c2 = .28
295
c1 = .29
150
c0 = .30
500
1,000
Q
The all-units schedule appears irrational in some respects. In Example 4.4, 499 bags
would cost $149.70, whereas 500 bags would cost only $145.00. Why would Weighty
actually charge less for a larger order? One reason would be to provide an incentive for
the purchaser to buy more. If you were considering buying 400 bags, you might choose
to move up to the breakpoint to obtain the discount. Furthermore, it is possible that
Weighty has stored its bags in lots of 100, so that its savings in handling costs might
more than compensate for the lower total cost.
Optimal Policy for All-Units Discount Schedule
We will illustrate the solution technique using Example 4.4. Assume that the company
considering what standing order to place with Weighty uses trash bags at a fairly constant
nah77856_ch04.indd 221
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 222
222
Chapter Four
Inventory Control
222 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
rate of 600 per year. The accounting department estimates that the fixed cost of placing
an order is $8, and holding costs are based on a 20 percent annual interest rate. From
Example 4.4, c0 0.30, c1 0.29, and c2 0.28 are the respective unit costs.
The first step toward finding a solution is to compute the EOQ values corresponding
to each of the unit costs, which we will label Q(0), Q(1), and Q(2), respectively.
Q(0) Q(1) Q(2) (2)(8)(600)
2Kl
400,
A Ic0
A (0.2)(0.30)
(2)(8)(600)
2Kl
406,
A Ic1
A (0.2)(0.29)
(2)(8)(600)
2Kl
414.
A Ic2
A (0.2)(0.28)
We say that the EOQ value is realizable if it falls within the interval that corresponds
to the unit cost used to compute it. Since 0 400 500, Q0 is realizable. However,
neither Q (1) nor Q (2) is realizable (Q (1) would have to have been between 500 and
1,000, and Q (2) would have to have been 1,000 or more). Each EOQ value corresponds
to the minimum of a different annual cost curve. In this example, if Q(2) were realizable, it would necessarily have to have been the optimal solution, as it corresponds to
the lowest point on the lowest curve. The three average annual cost curves for this example appear in Figure 4–11. Because each curve is valid only for certain values of Q,
the average annual cost function is given by the discontinuous curve shown in heavy
shading. The goal of the analysis is to find the minimum of this discontinuous curve.
There are three candidates for the optimal solution: 400, 500, and 1,000. In general,
the optimal solution will be either the largest realizable EOQ or one of the breakpoints
that exceeds it. The optimal solution is the lot size with the lowest average annual cost.
FIGURE 4–11
240
G(Q)
All-units discount
average annual cost
function
230
G0(Q )
220
G1(Q )
210
G2(Q )
200
190
180
100 200 300 400 500 600 700 800 900 1,000 1,100 1,200
Q
nah77856_ch04.indd 222
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 223
4.7 Quantity Discount
Discount Models 223
Models 223
4.7 Quantity
The average annual cost functions are given by
Gj (Q) cj KQ Icj Q2
for j 0, 1, and 2.
The broken curve pictured in Figure 4–11, G(Q), is defined as
G(Q) G0(Q)
G1(Q)
L G2(Q)
for
for
for
0 … Q 6 500,
500 … Q 6 1,000,
1,000 … Q
Substituting Q equals 400, 500, and 1,000, and using the appropriate values of cj,
we obtain
G(400) G0(400)
(600)(0.30) (600)(8)400 (0.2)(0.30)(400)2 $204.00
G(500) G1(500)
(600)(0.29) (600)(8)500 (0.2)(0.29)(500)2 $198.10
G(1,000) G2(1,000)
(600)(0.28) (600)(8)1,000 (0.2)(0.28)(1,000)2 $200.80.
Hence, we conclude that the optimal solution is to place a standing order for 500 units
with Weighty at an average annual cost of $198.10.
Summary of the Solution Technique for All-Units Discounts
1. Determine the largest realizable EOQ value. The most efficient way to do this is to
compute the EOQ for the lowest price first, and continue with the next higher price.
Stop when the first EOQ value is realizable (that is, within the correct interval).
2. Compare the value of the average annual cost at the largest realizable EOQ and at all
the price breakpoints that are greater than the largest realizable EOQ. The optimal Q
is the point at which the average annual cost is a minimum.
Incremental Quantity Discounts
Consider Example 4.4, but assume incremental quantity discounts. That is, the trash
bags cost 30 cents each for quantities of 500 or fewer; for quantities between 500 and
1,000, the first 500 cost 30 cents each and the remaining amount cost 29 cents each; for
quantities of 1,000 and over the first 500 cost 30 cents each, the next 500 cost 29 cents
each, and the remaining amount cost 28 cents each. We need to determine a mathematical expression for the function C(Q) pictured in Figure 4–10. From the figure, we see
that the first price break corresponds to C(Q) (500)(0.30) $150 and the second
price break corresponds to C(Q) 150 (0.29)(500) $295. It follows that
C(Q) so that
0.30Q
150 0.29(Q 500) 5 0.29Q
L 295 0.28(Q 1,000) 15 0.28Q
0.30
C(Q)>Q 0.29 5>Q
L 0.28 15>Q
nah77856_ch04.indd 223
for
for
for
for
for
for
0 … Q 6 500,
500 … Q 6 1,000,
1,000 … Q
0 … Q 6 500,
500 … Q 6 1,000,
1,000 … Q.
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 224
224
Chapter Four
Inventory Control
224 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
Average annual
cost function for
incremental discount
schedule
G(Q )
FIGURE 4–12
220
219
218
217
216
215
214
213
212
211
210
209
208
207
206
205
204
203
202
201
200
100 200 300 400 500 600 700 800 900 1,000 1,100 1,200
Q
The average annual cost function, G(Q), is
G(Q) C(Q)Q KQ I[C(Q)Q]Q2.
In this example, G(Q) will have three different algebraic representations [G0(Q), G1(Q),
and G2(Q)] depending upon into which interval Q falls. Because C(Q) is continuous,
G(Q) also will be continuous. The function G(Q) appears in Figure 4–12.
The optimal solution occurs at the minimum of one of the average annual cost
curves. The solution is obtained by substituting the three expressions for C(Q)Q in the
defining equation for G(Q), computing the three minima of the curves, determining
which of these minima fall into the correct interval, and, finally, comparing the average
annual costs at the realizable values. We have that
G0(Q) (600)(0.30) (8)(600)Q (0.20)(0.30)Q2
which is minimized at
Q(0) (2)(8)(600)
2Kl
400;
A Ic0
A (0.20)(0.30)
G1(Q) (600)(0.29 5Q) (8)(600)Q (0.20)(0.29 5Q)(Q2)
(0.29)(600) (13)(600)Q (0.20)(0.29)Q2 (0.20)(5)2
which is minimized at
Q(1) and finally
(2)(13)(600)
519;
A (0.20)(0.29)
G2(Q) (600)(0.28 15Q) (8)(600)Q (0.20)(0.28 15Q)Q2
(0.28)(600) (23)(600)Q (0.20)(0.28)Q2 (0.20)(15)2
nah77856_ch04.indd 224
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 225
4.7 Quantity Discount
Discount Models 225
Models 225
4.7 Quantity
which is minimized at
Q(2) (2)(23)(600)
702.
A (0.20)(0.28)
Both Q(0) and Q(1) are realizable. Q (2) is not realizable because Q (2) 1,000. The
optimal solution is obtained by comparing G0(Q (0)) and G1(Q (1)). Substituting into the
earlier expressions for G0(Q) and G1(Q), we obtain
G0(Q(0)) $204.00,
G1(Q(1)) $204.58.
Hence, the optimal solution is to place a standing order with the Weighty Trash Bag
Company for 400 units at the highest price of 30 cents per unit. The cost of using a standard order of 519 units is only slightly higher. Notice that compared to the all-units
case, we obtain a smaller batch size at a higher average annual cost.
Summary of the Solution Technique for Incremental Discounts
1. Determine an algebraic expression for C(Q) corresponding to each price interval.
Use that to determine an algebraic expression for C(Q)Q.
2. Substitute the expressions derived for C(Q)Q into the defining equation for G(Q).
Compute the minimum value of Q corresponding to each price interval separately.
3. Determine which minima computed in (2) are realizable (that is, fall into the correct
interval). Compare the values of the average annual costs at the realizable EOQ
values and pick the lowest.
Other Discount Schedules
Although all-units and incremental discount schedules are the most common, there are
a variety of other discount schedules as well. One example is the carload discount
schedule pictured in Figure 4–13.
Order cost function for
carload discount
schedule
C(Q)
FIGURE 4–13
Slope = c
Slope = c
M
M
Q
nah77856_ch04.indd 225
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 226
226
Chapter Four
Inventory Control
226 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
The rationale behind the carload discount schedule is the following. A carload consists of M units. The supplier charges a constant c per unit up until you have paid for
the cost of a full carload, at which point there is no charge for the remaining units in
that carload. Once the first carload is full, you again pay c per unit until the second
carload is full, and so forth.
Determining optimal policies for the carload or other discount schedules could be
extremely difficult. Each discount schedule has a unique solution procedure. Some can
be extremely complex.
Problems for Section 4.7
21. Your local grocery store stocks rolls of bathroom tissue in single packages and in
more economical 12-packs. You are trying to decide which to buy. The single
package costs 45 cents and the 12-pack costs $5. You consume bathroom tissue at
a fairly steady rate of one roll every three months. Your opportunity cost of money
is computed assuming an interest rate of 25 percent and a fixed cost of $1 for the
additional time it takes you to buy bathroom tissue when you go shopping. (We are
assuming that you shop often enough so that you don’t require a special trip when
you run out.)
a. How many single rolls should you be buying in order to minimize the annual
holding and setup costs of purchasing bathroom tissue?
b. Determine if it is more economical to purchase the bathroom tissue in 12-packs.
c. Are there reasons other than those discussed in the problem that would motivate you not to buy the bathroom tissue in 12-packs?
22. A purchasing agent for a particular type of silicon wafer used in the production of
semiconductors must decide among three sources. Source A will sell the silicon
wafers for $2.50 per wafer, independently of the number of wafers ordered. Source B
will sell the wafers for $2.40 each but will not consider an order for fewer than
3,000 wafers, and Source C will sell the wafers for $2.30 each but will not accept
an order for fewer than 4,000 wafers. Assume an order setup cost of $100 and an
annual requirement of 20,000 wafers. Assume a 20 percent annual interest rate for
holding cost calculations.
a. Which source should be used, and what is the size of the standing order?
b. What is the optimal value of the holding and setup costs for wafers when the
optimal source is used?
c. If the replenishment lead time for wafers is three months, determine the reorder
point based on the on-hand level of inventory of wafers.
23. Assume that two years have passed, and the purchasing agent mentioned in Problem
22 must recompute the optimal number of wafers to purchase and from which
source to purchase them. Source B has decided to accept any size offer, but sells the
wafers for $2.55 each for orders of up to 3,000 wafers and $2.25 each for the incremental amount ordered over 3,000 wafers. Source A still has the same price schedule, and Source C went out of business. Now which source should be used?
24. In the calculation of an optimal policy for an all-units discount schedule, you first
compute the EOQ values for each of the three order costs, and you obtain: Q (0) 800, Q (1) 875, and Q (2) 925. The all-units discount schedule has breakpoints
at 750 and 900. Based on this information only, can you determine what the optimal order quantity is? Explain your answer.
nah77856_ch04.indd 226
17/12/14 6:39 PM
Snapshot Application
SmartOps Assists in Designing
Caterpillar’s Inventory Control System
The Caterpillar Corporation is one of the world’s leading
manufacturers of construction and mining equipment,
diesel and natural gas engines, and gas turbines. Caterpillar posted annual sales of nearly $66 billion in 2012, and is
a Fortune 100 company. In 1997, management embarked
on a project aimed at improving product availability and
inventory turns. In 2001, the project team was expanded
to include both internal Caterpillar personnel and two
analysts from SmartOps, David Alberti and SmartOps
founder Sridhar Tayur. The team was charged with examining the following questions: (1) What product availability is possible at what cost and what inventory levels? (2)
How much inventory reduction is possible? (3) What mix
and deployment of inventory is necessary to achieve the
firm’s service objectives?
The team focused its attention on backhoe loaders
produced by the company’s Clayton, North Carolina,
­facility. A key part of their focus was on lead times.
These include both lead times for products being
shipped from the plant, and for materials and equipment shipped into the plant from suppliers. While lead
times coming out of the plant were relatively short
(typically one week), the lead times of products coming
from suppliers could be quite long, and also tended to
have high variance. The team developed a comprehensive model of the entire inventory supply chain. Inputs
to the model included inventory stocking locations, historical forecast errors for different seasons and d
­ ifferent
products, subassemblies and bill of materials for each
nah77856_ch04.qxd 1/9/08 3:21 PM Page 227
product, lead times, and inventory review times. The
model developed took into account the multi-echelon
(that is, multi-level) nature of the system.
In order to provide an accurate picture of the system, the analysts differentiated three types of orders
placed by dealers: (a) sold orders based on firm sales
to customers, (b) orders by dealers placed for the purpose of replenishing inventory, and (c) orders placed by
dealers to replenish stock of machines rented to customers. Note that replenishment lead times for each of
these demand types are different: four to six weeks for
sold orders, six weeks for inventory replenishment, and
fourteen weeks for replenishment of rental equipment.
The model provided a comprehensive picture of the
important service versus inventory trade-offs Caterpillar could expect from the Clayton plant. For example,
the team found that the Clayton plant could achieve a
consistent 45 and 30 day product availability to dealers
with 50 percent less finished goods inventory. However,
to achieve a 14 or 10 day availability the finished goods
inventory would have to be increased by approximately
90 percent. They also found that current service levels
could be maintained with a total supply chain inventory
reduction of 30–50 percent by repositioning inventory
from finished goods to components.
After careful consideration of the trade-offs involved,
management made the following changes to the system:
(1) total inventory in the supply chain was reduced by
16 percent, (2) the mean time for orders was reduced by
20 percent and the variance reduced by 50 percent, and
(3) order fulfillment was increased 2 percent.
The analysis provided by the team allowed management to better understand that there were no simple
answers to their inventory management problems. The
Caterpillar supply chain was very complex with interaction among the multiple levels. Seeing what trade-offs
were possible allowed management to set priorities
which ultimately resulted in lower inventory costs and
higher levels of service to the customer.
Source: Keene, et al. “Caterpillar’s Building Construction Products Division Improves and Stabilizes Product
­Availability” Interfaces 36 (2006), pp. 283–295.
4.8
Resource-Constrained Multiple Product Systems 227
25. Parasol Systems sells motherboards for personal computers. For quantities up
through 25, the firm charges $350 per board; for quantities between 26 and 50,
it charges $315 for each board purchased beyond 25; and it charges $285 each
for the additional quantities over 50. A large communications firm expects to
require these motherboards for the next 10 years at a rate of at least 140 per
year. Order setup costs are $30 and holding costs are based on an 18 percent
annual interest rate. What should be the size of the standing order?
*4.8
RESOURCE-CONSTRAINED MULTIPLE PRODUCT SYSTEMS
The EOQ model and its extensions apply only to single inventory items. However, these
models are often used in companies stocking many different items. Although we could
certainly compute optimal order quantities separately for each of the different items,
there could exist constraints that would make the resulting solution infeasible. In
Example 4.1, the optimal solution called for purchasing 3,870 pencils every 1.24 years.
The bookstore, however, may not have allocated enough space to store that many pencils,
nor enough money to purchase that many at one time.
Example 4.5
nah77856_ch04.indd 227
Three items are produced in a small fabrication shop. The shop management has established
the requirement that the shop never have more than $30,000 invested in the inventory of these
17/12/14
6:39 PM
year. Order setup costs are $30 and holding costs are based on an 18 percent
annual interest rate. What should be the size of the standing order?
*4.8
RESOURCE-CONSTRAINED MULTIPLE PRODUCT SYSTEMS
4.8
Resource-Constrained Multiple Product Systems 227
The EOQ model and its extensions apply only to single inventory items. However, these
models
are
often
used in
companies stocking many different items. Although we could
228 Chapter Four Inventory
Control
­Subject
to Known
­Demand
25. Parasol
Systems
sells order
motherboards
personalfor
computers.
Fordifferent
quantities
up
certainly
compute
optimal
quantitiesfor
separately
each of the
items,
through
25,
the
firm
charges
$350
per
board;
for
quantities
between
26
and
50,
there could exist constraints that would make the resulting solution infeasible. In
it charges
$315
for each
board
purchased
beyond3,870
25; and
it charges
$285years.
each
Example
4.1, the
optimal
solution
called
for purchasing
pencils
every 1.24
for
the
additional
quantities
over
50.
A
large
communications
firm
expects
to
The bookstore, however, may not have allocated enough space to store that many pencils,
require
these
motherboards
for
the
next
10
years
at
a
rate
of
at
least
140
per
nor enough money to purchase that many at one time.
year. Order setup costs are $30 and holding costs are based on an 18 percent
annual interest rate. What should be the size of the standing order?
Example 4.5
Three items are produced in a small fabrication shop. The shop management has established
*4.8
the requirement that the shop never have more than $30,000 invested in the inventory of these
items at one time. The management
uses a 25 percent
annual interest
charge to compute the
RESOURCE-CONSTRAINED
MULTIPLE
PRODUCT
SYSTEMS
holding cost. The relevant cost and demand parameters are given in the following table. What
lot sizes should the shop be producing so that they do not exceed the budget?
The EOQ model and its extensions apply only to single inventory items. However, these
models are often used in companies stocking many different items. Although we could
Itemfor each of the different items,
certainly compute optimal order quantities separately
there could exist constraints that would
infeasible. In
1 make the 2resulting solution
3
Example 4.1, the optimal
solution
called
for
purchasing
3,870
pencils
every
1.24 years.
Demand rate j
1,850
1,150
800
The bookstore, however,
may
not
have
allocated
enough
space
to
store
that
many
pencils,
Variable cost cj
50
350
85
nor enough moneySetup
to purchase
that
many
at
one
time.
cost Kj
100
150
50
Example 4.5
Solution
nah77856_ch04.qxd
1/9/08
Three items are produced in a small fabrication shop. The shop management has established
If
the
budget is not
the EOQ
of these
three
items,
then the
the
requirement
thatexceeded
the shop when
never using
have more
thanvalues
$30,000
invested
in the
inventory
of EOQs
these
are
optimal.
Hence,
the
first
step
is
to
compute
the
EOQ
values
for
all
items
to
determine
items at one time. The management uses a 25 percent annual interest charge to compute the
whether
or not
the
constraint
active.
holding cost.
The
relevant
costisand
demand parameters are given in the following table. What
lot sizes should the shop be producing so that they do not exceed the budget?
1221100211,8502
172,
EOQ1 A 10.2521502
Item
1221150211,1502
1
2 63,
3
EOQ2 A 10.25213502
3:21 PM Page 228 Demand rate j
1,850
1,150
800
Variable cost cj
50
350
85
122150218002
150
61.
Setup costEOQ
Kj 3 A 100
50
10.2521852
If the EOQ value for each item is used, the maximum investment in inventory would be
If the budget is not exceeded when using the EOQ values of these three items, then the EOQs
3
(61)(85)
$35,835.
are optimal. Hence, the(172)(50)
first stepis(63)(350)
to compute
the EOQ
values for
all items to determine
228 Chapter Four Inventory Control Subject to Known Demand
whether or not the constraint is active.
Because the EOQ solution violates the constraint, we need to reduce these lot sizes. But how?
Solution
The optimal solution turns out to be very
easy to find in this case. We merely multiply each
1221100211,8502
We are assuming for the purposes
of1this
products
section that the three 172, are ordered simultaneously. By
EOQ
EOQ value by the ratio (30,000)(35,835)
�
0.8372. In order to guarantee that we do not
A
staggering order cycles, it is possible to meet the 10.2521502
constraint with larger lot sizes. We will not consider that
exceed
the $30,000 budget, we round each value down (adjustments can be made subsecase
here.
quently). Letting Q 1*, Q 2*, and Q 3* be the1221150211,1502
optimal values, we obtain
EOQ2 63,
A 10.25213502
Q 1* � (172)(0.8372) 144,
3
122150218002
*
61.
EOQQ32� (63)(0.8372) 52,
A
10.2521852 51,
Q 3* � (61)(0.8372)
If the
value
for each item
used, the
investment
where
EOQ
should
be interpreted
as is
rounding
tomaximum
the next lower
integerinininventory
this case.would be
3
The total budget required
for these
lot sizes
$29,735.
The remaining
$265 can now be
(172)(50)
(63)(350)
is
(61)(85)
$35,835.
used to increase the lot sizes of products 1 and 3 slightly. (For example, Q 1* � 147, Q 3* � 52
Because
EOQ solution
violates the constraint, we need to reduce these lot sizes. But how?
results inthe
a budget
of $29,970.)
3
We
assuming
for the purposes
of this section that
the three products
are ordered
simultaneously.
By
Inare
general,
budgetor space-constrained
problems
are not solved
so easily.
Suppose
staggering
orderhave
cycles,
it is costs
possible
the
constraint
with larger lotand
sizes.
Wetotal
will not
consider
that
,
c
,
.
.
.
,
c
,
respectively,
the
budget
availthat n items
unit
oftocmeet
1 2
n
case here.
able for them is C. Then the budget constraint can be written
c1Q1 � c2Q2 � � � � � cnQn � C.
Let
EOQi �
nah77856_ch04.indd 228
2Kili
A hi
for i � 1, . . . , n,
where Ki, hi, and i are the respective cost and demand parameters.
There are two possibilities: either the constraint is active or it is not. If the constraint
is not active, then
17/12/14 6:39 PM
Q 1* � (172)(0.8372) 144,
Q 2* � (63)(0.8372) 52,
Q 3* � (61)(0.8372) 51,
where should be interpreted as rounding to the next lower integer in this case.
The total budget required for these lot sizes is $29,735. The remaining $265 can now be
used to increase the lot sizes of products 1 and 3 slightly. (For example, Q 1* � 147, Q 3* � 52
4.8 Resource-Constrained Multiple Product Systems 229
results in a budget of $29,970.)
In general, budget- or space-constrained problems are not solved so easily. Suppose
that n items have unit costs of c1, c2, . . . , cn, respectively, and the total budget available for them is C. Then the budget constraint can be written
c1Q1 � c2Q2 � � � � � cnQn � C.
Let
EOQi �
2Kili
A hi
for i � 1, . . . , n,
where Ki, hi, and i are the respective cost and demand parameters.
There are two possibilities: either the constraint is active or it is not. If the constraint
is not active, then
a ciEOQi … C,
n
i�1
and the optimal solution is Qi � EOQi. If the constraint is active, then
a ciEOQi 7 C,
n
i�1
and the EOQ solution is no longer feasible. If we include the following assumption, the
solution to the active case is relatively easy to find:
Assumption: c1h1 � c2h2 � � � � � cnhn .
If this assumption holds and the constraint is active, we prove in Appendix 4–A that
the optimal solution is
Q*i � mEOQi ,
where the multiplier m solves
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 229
m � C n c a (ciEOQi) d .
n
i�1
Since cihi � ci(Iici) � 1Ii, the condition that the ratios be equal is equivalent to
the requirement that the same interest rate be used to compute the holding cost for each
item, which is reasonable in most circumstances.
4.8 Resource-Constrained
229
be the Product
spaceSystems
consumed
Suppose that the constraint is on the available
space. Let wi Multiple
by one unit of product i for i � 1, 2, . . . , n (this could be floor space measured, say, in
square feet, or volume measured in cubic feet), and let W be the total space available.
Then the space constraint is of the form
w1Q1 � w2Q2 � � � � � wnQn � W.
This is mathematically of the same form as the budget constraint, so the same analysis
applies. However, our condition for a simple solution now is that the ratios wi hi are
equal. That is, the space consumed by an item should be proportional to its holding cost.
When the interest rate is fixed, this is equivalent to the requirement that the space consumed should be proportional to the value of the item. This requirement would probably be too restrictive in most cases. For example, fountain pens take up far less space
than legal pads, but are more expensive.
Let us now consider the problem in which the constraint is active, but the proportionality assumption is not met. This problem is far more complex than that just solved. It requires formulating the Lagrangian function. The details of the formulation of this problem can be found in Appendix 4–A. As we show, the optimal lot sizes are now of the form
Q*i �
where � is a constant chosen so that
nah77856_ch04.indd 229
n
2Kili
,
A hi � 2uwi
w Q* � W.
17/12/14 6:39 PM
w1Q1 � w2Q2 � � � � � wnQn � W.
This is mathematically of the same form as the budget constraint, so the same analysis
applies. However, our condition for a simple solution now is that the ratios wi hi are
equal. That is, the space consumed by an item should be proportional to its holding cost.
When the interest rate is fixed, this is equivalent to the requirement that the space consumed should be proportional to the value of the item. This requirement would prob230 Chapter Four Inventory
­Subject
to Knownin
­Demand
ablyControl
be too
restrictive
most cases. For example, fountain pens take up far less space
than legal pads, but are more expensive.
Let us now consider the problem in which the constraint is active, but the proportionality assumption is not met. This problem is far more complex than that just solved. It requires formulating the Lagrangian function. The details of the formulation of this problem can be found in Appendix 4–A. As we show, the optimal lot sizes are now of the form
Q*i �
where � is a constant chosen so that
2Kili
,
A hi � 2uwi
*
a wiQi � W.
n
i�1
The constant �, known as the Lagrange multiplier, reduces the lot sizes by increasing
the effective holding cost. The correct value of � can be found by trial and error or by a
search technique such as interval bisection. Note that � � 0, so that the search can be
limited to positive numbers only. The value of � can be interpreted as the decrease in the
average annual cost that would result from adding an additional unit of resource. In this
case, it would represent the marginal benefit of an additional square foot of space.
Example 4.6
Consider the fabrication shop of Example 4.5. In addition to the budget constraint, suppose
that there are only 2,000 square feet of floor space available. Assume that the three products
consume respectively 9, 12, and 18 square feet per unit.
First, we check to see if the EOQ solution is feasible. Setting w1 � 9, w2 � 12, and w3 � 18,
we find that
a EOQi wi � (172)(9) � (63)(12) � (61)(18) � 3,402,
nah77856_ch04.qxd
1/9/08
which is obviously infeasible. Next, we check if the budget-constrained solution provides a
feasible solution to the space-constrained problem. The budget-constrained solution requires
(147)(9) � (52)(12) � (52)(18) � 2,883 square feet of space, which is also infeasible.
The next step is to compute the ratios wihi for 1 � i � 3. These ratios turn out to be 0.72,
0.14, and 0.85, respectively. Because their values are different, the simple solution obtained by
a proportional scaling of the EOQ values will not be optimal. Hence, we must determine the
value of the Lagrange multiplier �.
3:21 PM Page 230
We can determine upper and lower bounds on the optimal value of � by assuming equal
ratios. If the ratios were equal, the multiplier, m, would be
m � W a (EOQiwi) � 2,0003,402 � 0.5879,
which would give the three lot sizes as 101, 37, and 36, respectively. The three values of � that
result in these lot sizes are respectively � � 1.32, � � 6.86, and � � 2.33. (These values were
230 Chapter Four Inventory Control Subject to Known Demand
obtained by setting the given expression for Q1* equal to the lot sizes 101, 37, and 36, and
solving for �.)
The true value of � will be between 1.32 and 6.86. If we start with � � 3.5, we obtain
Q 1* � 70, Q 2* � 45, and Q *
3 � 23, and wi Q*
i � (70)(9) � (45)(12) � (23)(18) � 1,584, which
implies � � 3.5. After considerable experimentation, we finally find that the optimal value of
� � 1.75, and Q 1* � 92, Q 2* � 51, and Q 3* � 31, giving wiQ*i � 1,998. Notice that these lot
sizes are in very different proportions from the ones obtained assuming a constant multiplier.
Searching for the optimal value of the Lagrange multiplier, although tedious to do by hand,
can be accomplished very efficiently using a computer. Spreadsheets are a useful means for
effecting such calculations.
We have only touched on some of the complications that could arise when applying
EOQ-type models to a realistic system with multiple products. Often, real problems are
far too complex to be expressed accurately as a solvable mathematical model. For this
reason, simple models such as the ones presented in this and subsequent chapters are
used by practitioners. However, any solution recommended by a mathematical model
must be considered in the context of the system in which it is to be used.
Problems for Section 4.8
nah77856_ch04.indd 230
26. A local outdoor vegetable stand has exactly 1,000 square feet of space to display
three vegetables: tomatoes, lettuce, and zucchini. The appropriate data for these
17/12/14 6:39 PM
The true value of � will be between 1.32 and 6.86. If we start with � � 3.5, we obtain
Q 1* � 70, Q 2* � 45, and Q *
3 � 23, and wi Q*
i � (70)(9) � (45)(12) � (23)(18) � 1,584, which
implies � � 3.5. After considerable experimentation, we finally find that the optimal value of
� � 1.75, and Q 1* � 92, Q 2* � 51, and Q 3* � 31, giving wiQ*i � 1,998. Notice that these lot
sizes are in very different proportions from the ones obtained assuming a constant multiplier.
Searching for the optimal value of the Lagrange multiplier, although tedious to do by hand,
can be accomplished very efficiently using a computer. Spreadsheets are a useful means for
effecting such calculations.
4.9 EOQ Models for Production Planning 231
We have only touched on some of the complications that could arise when applying
EOQ-type models to a realistic system with multiple products. Often, real problems are
far too complex to be expressed accurately as a solvable mathematical model. For this
reason, simple models such as the ones presented in this and subsequent chapters are
used by practitioners. However, any solution recommended by a mathematical model
must be considered in the context of the system in which it is to be used.
Problems for Section 4.8
26. A local outdoor vegetable stand has exactly 1,000 square feet of space to display
three vegetables: tomatoes, lettuce, and zucchini. The appropriate data for these
items are given in the following table.
Item
Tomatoes
Annual demand
(in pounds)
Cost per pound
nah77856_ch04.qxd
4.9
1/9/08
3:21
Lettuce
Zucchini
850
1,280
630
$0.29
$0.45
$0.25
The setup cost for replenishment of the vegetables is $100 in each case, and the
space consumed by each vegetable is proportional to its costs, with tomatoes requiring 0.5 square foot per pound. The annual interest rate used for computing
holding costs is 25 percent. What are the optimal quantities that should be purchased of these three vegetables?
27. Suppose that the vegetables discussed in Problem 26 are purchased at different
times. In what way could that have an effect on the order policy that the stand owner
should use?
28. Suppose that in Problem 26 the space consumed by each vegetable is not
proportional to its cost. In particular, suppose that one pound of lettuce required
0.4 square foot of space and one pound of zucchini required 1 square foot of space.
Determine upper and lower bounds on the optimal values of the order quantities in
this case. Test different values of the Lagrange multiplier to find the optimal values of the order quantities. (A spreadsheet is ideally suited for this kind of calculation. If you solve the problem using a spreadsheet, place the Lagrange multiplier
PM Page
in a231
cell so that its value can be changed easily.)
EOQ MODELS FOR PRODUCTION PLANNING
Simple lot-sizing models have been successfully applied to a variety of manufacturing
4.9 of
EOQ
Models
for Production
Planning
231
problems. In this section we consider an extension
the
EOQ
model with
a finite
production rate, discussed in Section 4.6, to the problem of producing n products on a
single machine. Following the notation used in this chapter, let
j � Demand rate for product j,
Pj � Production rate for product j,
hj � Holding cost per unit per unit time for product j,
Kj � Cost of setting up the production facility to produce product j.
The goal is to determine the optimal procedure for producing n products on the machine to minimize the cost of holding and setups, and to guarantee that no stock-outs
occur during the production cycle.
We require the assumption that
a lj>Pj … 1.
n
j�1
nah77856_ch04.indd 231
This assumption is needed to ensure that the facility has sufficient capacity to satisfy
the demand for all products. Notice that this is stronger than the assumption made in
17/12/14 6:39 PM
single machine. Following the notation used in this chapter, let
j � Demand rate for product j,
Pj � Production rate for product j,
hj � Holding cost per unit per unit time for product j,
232 Chapter Four Inventory Control ­Subject to Known ­Demand
Kj � Cost of setting up the production facility to produce
product j.
The goal is to determine the optimal procedure for producing n products on the machine to minimize the cost of holding and setups, and to guarantee that no stock-outs
occur during the production cycle.
We require the assumption that
a lj>Pj … 1.
n
j�1
This assumption is needed to ensure that the facility has sufficient capacity to satisfy
the demand for all products. Notice that this is stronger than the assumption made in
Section 4.6 that j � Pj for each j. To see why this assumption is necessary, consider the
case of two products with identical demand and production rates. In each cycle one
would produce product 1 first, then product 2. Clearly P1 � 21 and P2 � 22, so that
enough could be produced to meet the total demand for both products each cycle. This
translates to 1P1 � 0.5 and 2P2 � 0.5, giving a value of the sum less than or equal
to one. Similar reasoning holds for more than two products with nonidentical demand
and production rates.
We also will assume that the policy used is a rotation cycle policy. That means that
in each cycle there is exactly one setup for each product, and products are produced in
the same sequence in each production cycle. The importance of this assumption will be
discussed further at the end of the section.
At first, one might think that the optimal solution is to sequentially produce lot sizes
for each product optimized by treating each product in isolation. From Section 4.6, this
would result in a lot size for product j of
Qj �
2Kjlj
,
A h¿j
where h�j � hj (1 � jPj). The problem with this approach is that because we have only
a single production facility, it is likely that some of the lot sizes Qj will not be large
enough to meet the demand between production runs for product j, thus resulting in
stock-outs.
Let T be the cycle time. During time T we assume that exactly one lot of each product is produced. In order that the lot for product j will be large enough to meet the
demand occurring during time T, it follows that the lot size must be
Qj � jT.
nah77856_ch04.qxd
1/9/08
Section
3:21 From
PM Page
232 4.6,
the average annual cost associated with product j can be written
in the form
G(Qj) � Kj jQj � h�j Qj2.
The average annual cost for all products is the sum
232 Chapter Four
Inventory Control Subject to Known Demand
a G(Qj) � a Kjlj>Qj � hj¿Qj>2.
n
n
j�1
j�1
Substituting T � Qjj, we obtain the average annual cost associated with the n
products in terms of the cycle time T as
G(T) � a [Kj>T � hj¿ljT>2].
n
j�1
The goal is to find T to minimize G(T). The necessary condition for an optimal T is
dG(T)
� 0.
dT
Setting the first derivative with respect to T to zero gives
2
a [�Kj>T � hj¿lj>2] � 0.
n
nah77856_ch04.indd 232
j�1
17/12/14 6:39 PM
232 Chapter Four Inventory Control Subject to Known Demand
Substituting T � Qjj, we obtain the average annual cost associated with the n
products in terms of the cycle time T as
G(T) � a [Kj>T � hj¿ljT>2].
4.9 EOQ Models for Production Planning 233
n
j�1
The goal is to find T to minimize G(T). The necessary condition for an optimal T is
dG(T)
� 0.
dT
Setting the first derivative with respect to T to zero gives
2
a [�Kj>T � hj¿lj>2] � 0.
n
j�1
Solving for T, we obtain the optimal cycle time T* as
2 a Kj
n
T* �
j�1
n
a
d j�1
.
hj¿lj
If setup times are a factor, we must check that there is enough time each cycle to account for both setup times and production of the n products. Let sj be the setup time for
product j. Ensuring that the total time required for setups and production each cycle
does not exceed T leads to the constraint
a (sj � Qj>Pj) … T.
n
j�1
Using the fact that Qj � jT, this condition translates to
a (sj � ljT>Pj) … T,
n
j�1
which gives, after rearranging terms,
a sj
n
T Ú
j�1
n
1 � a (lj>Pj)
� Tmin.
j�1
Because Tmin cannot be exceeded without compromising feasibility, the optimal
solution is to choose the cycle time T equal to the larger of T* and Tmin.
Example 4.7
nah77856_ch04.indd 233
Bali produces several styles of men’s and women’s shoes at a single facility near Bergamo, Italy.
The leather for both the uppers and the soles of the shoes is cut on a single machine. This
Bergamo plant is responsible for seven styles and several colors in each style. (The colors are not
considered different products for our purposes, because no setup is required when switching
colors.) Bali would like to schedule cutting for the shoes using a rotation policy that meets all
demand and minimizes setup and holding costs. Setup costs are proportional to setup times.
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 233
234 Chapter Four Inventory Control ­Subject to Known ­Demand
TABLE 4–1
Relevant Data for
Example 4.7
Style
4.9
EOQ Models for Production Planning 233
Annual Demand
(units/year)
Production Rate
(units/year)
Setup Time
(hours)
Variable Cost
($/unit)
4,520
6,600
2,340
2,600
8,800
6,200
5,200
35,800
62,600
41,000
71,000
46,800
71,200
56,000
3.2
2.5
4.4
1.8
5.1
3.1
4.4
$40
26
52
18
38
28
31
Women’s pump
Women’s loafer
Women’s boot
Women’s sandal
Men’s wingtip
Men’s loafer
Men’s oxford
The firm estimates that setup costs amount to an average of $110 per hour, based on the cost
of worker time and the cost of forced machine idle time during setups. Holding costs are based
on a 22 percent annual interest charge.
The relevant data for this problem appear in Table 4–1.
Solution
The first step is to verify that the problem is feasible. To do so we compute jPj. The reader
should verify that this sum is equal to 0.69355. Because this is less than one, there will be a feasible solution. Next we compute the value of T*, but to do so we need to do several intermediate calculations.
First, we compute setup costs. Setup costs are assumed to be $110 times setup times. Second,
we compute modified holding costs (hj ). This is done by multiplying the cost of each product by
the annual interest rate (0.22) times the factor 1 jPj. These calculations give
Setup Costs (Kj)
352
275
484
198
561
341
484
Modified Holding Costs (hj )
7.69
5.12
10.79
3.81
6.79
5.62
6.19
The sum of the setup costs is 2,695, and the sum of the products of the modified holding
costs and the annual demands is 230,458.4. Substituting these figures into the formula for T*
gives the optimal cycle time as 0.1529 year. Assuming a 250-day work year, this means that the
rotation cycle should repeat roughly every 38 working days. The optimal lot size for each of the
shoes is found by multiplying the cycle time by the demand rate for each item. The reader
should check that the following lot sizes result:
Style
Women’s pump
Women’s loafer
Women’s boot
Women’s sandal
Men’s wingtip
Men’s loafer
Men’s oxford
Optimal Lot Sizes for
Each Production Run
691
1,009
358
398
1,346
948
795
The plant would cut the soles and uppers in these lot sizes in sequence (although the
sequence does not necessarily have to be this one) and would repeat the rotation cycle roughly
nah77856_ch04.indd 234
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 234
234 Chapter Four Inventory Control Subject to Known Demand
4.9 EOQ Models for Production Planning 235
every 38 days (0.1529 year). However, this solution can be implemented only if T* is at least
Tmin. To determine Tmin we must express the setup times in years. Assuming 8 working hours
per day and 250 working days per year, one would divide the setup times given in hours by
2,000 (250 times 8). The reader should check that the resulting value of Tmin is 0.04, thus making T* feasible and, hence, optimal.
The total average annual cost of holding and setups at an optimal policy can be found by
computing the value of G(T) when T � T*. It is $35,244.44. It is interesting to note that if the
plant manager chooses to implement this policy, the facility will be idle for a substantial portion
of each rotation cycle. The total uptime each rotation cycle is found by dividing the lot sizes by
the production rates for each style and summing the results. It turns out to be 0.106 year.
Hence, the optimal rotation policy that minimizes total holding and setup costs results in the
cutting operation remaining idle about one-third of the time.
The reader should be aware that the relatively simple solution to this problem was the
result of two assumptions. One was that the setup costs were not sequence dependent. In
Example 4.7, it is possible that the time required to change over from one shoe style to
another could depend on the styles. For example, a changeover from a woman’s style to
a man’s style probably takes longer than from one woman’s style to another. A second assumption was that the plant used a rotation cycle policy. That is, in each cycle Bali does
a single production run of each style. When demand rates and setup costs differ widely,
it might be advantageous to do two or more production runs of a product in a cycle. This
more general problem has not, to our knowledge, been solved. A discussion of these and
related issues can be found in Maxwell (1964). Magee and Boodman (1967) provide
some heuristic ways of dealing with the more general problem.
Problems for Section 4.9
29. A metal fabrication shop has a single punch press. There are currently three parts
that the shop has agreed to produce that require the press, and it appears that they
will be supplying these parts well into the future. You may assume that the press is
the critical resource for these parts, so that we need not worry about the interaction
of the press with the other machines in the shop. The relevant information here is
Part
Number
1
2
3
Annual Contracted
Amount (demand)
Setup Cost
Cost
(per unit)
Production
Rate (per year)
2,500
5,500
1,450
$80
120
60
$16
18
22
45,000
40,000
26,000
Holding costs are based on an 18 percent annual interest rate, and the products
are to be produced in sequence on a rotation cycle. Setup times can be considered
negligible.
a. What is the optimal time between setups for part number 1?
b. What percentage of the time is the punch press idle, assuming an optimal
rotation cycle policy?
c. What are the optimal lot sizes of each part put through the press at an optimal
solution?
d. What is the total annual cost of holding and setup for these items on the punch
press, assuming an optimal rotation cycle?
nah77856_ch04.indd 235
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 235
4.10
236 Chapter Four Inventory Control ­Subject to Known ­Demand
Power-of-Two Policies 235
30. Tomlinson Furniture has a single lathe for turning the wood for various furniture
pieces, including bedposts, rounded table legs, and other items. Four forms are
turned on the lathe and produced in lots for inventory. To simplify scheduling, one lot
of each type will be produced in a cycle, which may include idle time. The four products and the relevant information concerning them appears in the following table.
Piece
Monthly
Requirements
Setup
Time (hours)
Unit
Cost
Production
Rate (units/day)
J–55R
H–223
K–18R
Z–344
125
140
45
240
1.2
0.8
2.2
3.1
$20
35
12
45
140
220
100
165
Worker time for setups is valued at $85 per hour, and holding costs are based on
a 20 percent annual interest charge. Assume 20 working days per month and
12 months per year for your calculations.
a. Determine the optimal length of the rotation cycle.
b. What are the optimal lot sizes for each product?
c. What are the percentages of uptime and downtime for the lathe, assuming that it
is not used for any other purpose?
d. Draw a graph showing the change in the inventory level over a typical cycle for
each product.
e. Discuss why the solution you obtained might not be feasible for the firm, or why
it might not be desirable even when it is feasible.
4.10
POWER-OF-TWO POLICIES
The inventory models treated in this chapter form the basis of more complex cases. In
almost every case treated here, we were able to find relatively straightforward algebraic
solutions. However, even when demand is assumed known, there exist several extensions of these basic models whose optimal solutions may be difficult or impossible to
find. In those cases, effective approximations are very valuable. Here, we discuss an
approach that has proven to be successful in a variety of deterministic environments.
The idea is based on choosing the best replenishment interval from a set of possible intervals proportional to powers of two. While the analysis of power-of-two policies in
complex environments is beyond the scope of this book, we can illustrate the idea
in the context of the basic EOQ model.
From Section 4.5, we know that the order quantity that minimizes average annual holding and setup costs when demand is fixed at units per unit time is the EOQ given by
2Kl
,
A h
and the optimal time between placement of orders, say T*, is given by
Q* T* Q*>l nah77856_ch04.indd 236
2K
.
A lh
17/12/14 6:39 PM
nah77856_ch04.qxd
236
1/9/08
3:21 PM
Page 236
Chapter Four Inventory Control Subject to Known Demand
4.10 Power-of-Two Policies 237
It is possible, and even likely, that optimal order intervals are inconvenient. For example, the optimal solution might call for ordering every 3.393 weeks. However, one
might only want to place orders at the beginning of a day or a week. To account for this,
suppose that we impose the constraint that ordering must occur in some multiple of a
base time, TL. To find the optimal solution under the constraint that the order interval
must be a multiple of TL, one would simply compare the costs at the two closest multiples of TL to T* and pick the order interval with the lower cost. [That is, find k for
which kTL … * … (k � 1)TL and choose the reorder interval to be either kTL or
(k � 1)TL depending on which results in a lower average annual cost.]
Now suppose we add the additional restriction that the order interval must be a
power of two times TL. That is, the order interval must be of the form 2kTL for some
integer k Ú 0. While it is unlikely one would impose such a restriction in the context
of the simple EOQ problem, the ultimate goal is to explore these policies for more
complex problems whose optimal solutions are hard to find. The question we wish to
address is: Under such a restriction (known as a power-of-two policy), what is the
worst cost error we will incur relative to that of the optimal reorder interval T*? On the
surface, it appears that such a restriction would result in serious errors. As k increases,
the distance between successive powers of two grows rapidly. For example, if k � 12,
2k � 4,096 and 2k�1 � 8,192, a very wide interval. If T* � 6,000 and TL � 1, for example, it seems that forcing the order interval to be either 4,096 or 8,192 would result
in a large cost error (as the error in T is nearly 30 percent in either direction). However,
this turns out not to be the case. In fact, we can prove that the cost error in every case
is bounded by slightly more than 6 percent. While the result seems unintuitive at first,
recall that the average annual cost function is relatively insensitive to errors in Q, as we
saw in Section 4.5. Since Q and T are proportional, a similar cost insensitivity holds
with respect to T.
We know from Section 4.5 that the average annual holding and setup cost as a function of the order quantity, Q, is given by
G(Q) �
Kl hQ
�
.
Q
2
Since Q � T, the average annual cost can also be expressed in terms of the reorder
interval, T, as
G(T) �
K hlT
�
.
T
2
In the case of a power-of-two policy, we wish to find the value of k that minimizes
G(2kTL). Since G(T ) is a continuous convex function of T, it follows that G(2kTL) is a
discrete convex function of k. That means that the optimal value of k, say k*, satisfies
k* � min{k :G (2k�1TL) Ú G (2kTL)}.
Substituting for G(T), the optimality condition becomes
K
2k�1TL
�
K
2kTL
Ú
hl2kTL hl2k�1TL
�
2
2
which can easily be seen to reduce to
K>2
2kTL
nah77856_ch04.indd 237
… hl2k–1TL.
17/12/14 6:39 PM
nah77856_ch04.qxd
1/9/08
3:21 PM
Page 237
4.11
238 Chapter Four Inventory Control ­Subject to Known ­Demand
Historical Notes and Additional Topics 237
Rearranging terms gives
2k Ú
or
1
K
,
TL A hl
2kTL Ú
1
22
T*.
We assume that T* Ú TL, which means that 22T* 7 TL. Hence, to summarize, we
seek the smallest value of k that satisfies the simultaneous inequalities
1
T* … 2kTL … 22T *.
22
Note that this implies that as long as T* Ú TL, the optimal power-of-two solution will
always lie between .707T* and 1.41T*. The next question is: What is the worst-case
cost error of this policy? Since Q lT, if T T*> 22, then Q Q*> 22, and similarly if T 22T*, then Q 22Q*. It follows that the worst-case cost error of the
power-of-two policy is given by
Q
G(Q)
1 Q*
1 1
a
b a
22b 1.0607,
G(Q*) 2 Q
Q*
2 22
or slightly more than 6 percent. (Because of symmetry, we obtain the same result when
Q Q*> 22 or when Q 22Q*.)
The real “power” of power-of-two policies occurs when trying to solve more complex problems whose optimal policies are difficult to find. Consider the following
scenario. A single warehouse is the sole supplier of N retailers. The demand rate experienced by each retailer is known and constant. As with the simple EOQ problem,
assume that shortages at both the warehouse and the retailers are not permitted. There
are fixed costs for ordering at both the warehouse and the retailers, and holding costs
at these locations. These costs do not necessarily need to be the same. It is assumed that
there is no lead time for placement or arrival of orders.
Unlike the simple EOQ problem, determining an optimal policy (that is, one that
minimizes long-run average costs) for this problem could be extremely difficult, or even
close to impossible. In many cases, even the form of an optimal policy may not be
known. Clearly, effective approximations are very important.
One approximation for this problem is a so-called nested policy. In this case, a retailer
would automatically order whenever the warehouse does, and possibly at other times as
well. Although nested policies can have arbitrarily large cost errors, various adaptations
of power-of-two policies can be shown to have 94 percent or even 98 percent guaranteed
effectiveness. Power-of-two approximations have also been shown to be equally effective in serial production systems, and more complex arborescent assembly systems. We
refer the interested reader to Roundy (1985) and Muckstadt and Roundy (1993).
Case Study. Betty Buys a Business
4.11
Betty Robinson decided it was time for a change. She had been a social worker for
HISTORICAL
NOTES
AND
TOPICS
10 years,
and although
sheADDITIONAL
found the work rewarding,
it was time for something else.
So, when Herbie’s Hut came up for sale, she made some inquiries.
TheHerbie’s
interestHut
in using
models
to store
control
the old
replenishment
of inventories
was amathematical
small gift shop
and toy
in her
neighborhood
of Skokie,
dates
back
to
the
early
part
of
the
20th
century.
Ford
Harris
(1915)
is
generally
a Chicago suburb. Herb Gold had been running the business for 40 years, butcredited
it was
the development
simpleitEOQ
H. Wilson
(1934)
is also recognized
awith
demanding
job and of
hethe
decided
was model.
time toR.hang
it up. The
business
was pretty
steady, so Herb figured that there would buyers. He was delighted to hear from Betty,
nah77856_ch04.indd 238
17/12/14 6:39 PM
4.10 Power-of-Two Policies 239
whom he had known as a child growing up in the neighborhood. The money was less
important to Herb than making sure the store was in good hands. He had a good feeling
about Betty, so he agreed to let the business go for a modest price.
After Betty purchased the business, she decided to remodel the store. Herbie’s Hut
would reopen as Betty’s Best. Skylights, new wallpaper, and a fresh layout gave the store
a sprightlier look that reflected Betty’s personality. During this time, Betty took a careful
look at the stock that came with the purchase. She noticed that several items were severely
understocked. Two of those were ID bracelets and Lego sets. She spoke with Herb about it
and he apologized. These were popular items and he hadn’t had a chance to reorder them.
Betty was concerned that she might run out of these items quickly and that it would
hurt her reputation. When she called the suppliers she realized that she had no idea how
much to order. She asked Herb and he said that his rule of thumb was to order about one
month’s supply. This sounded reasonable to Betty; she could figure out what a month’s
supply was from the store’s sales records, which Herb graciously turned over to her.
The store stocked hundreds of different items. It occurred to Betty that if she reordered every item monthly, she would be spending a lot of time processing orders. Betty
thought that perhaps there was a better way. She consulted her boyfriend Bob who had
an MBA degree. Bob fished up his old class notes and showed Betty the EOQ formula,
2Kl
, which was supposed to be used for determining order sizes. “What are these
Ä h
symbols supposed be?” asked Betty. Bob had reviewed his notes and explained that K
is the fixed cost of each order, a Greek letter (λ) is the sales rate, and h is the cost of
holding. Betty still wasn’t sure what it all meant.
“Let’s take an item you need to order now and see how this works out,” suggested
Bob. So they considered the bracelets.
“First, what’s the process you would go through to place a new order?” asked Bob. “Well,”
replied Betty, “I would call the supplier to place the order, and when it arrived I would unpack
the shipment, possibly put some of the items on display, and store the rest.” “How long does
that take?” inquired Bob. “I don’t know,” she responded, “maybe a couple of hours. Also,
that supplier charges a fixed cost of $50 per order in addition to the cost of each bracelet.”
Figuring Betty’s time at $50 per hour Bob computed a total fixed cost of (2)(50) + 50 = $150.
“Ok. Let’s look at the sales rate,” said Bob. Based on Herb’s records, she estimated that he
had sold an average of about 75 bracelets a month. That would give the value of λ. Finally,
they needed to estimate the holding cost. Bob had spent some time thinking about this. “Holding cost can be thought of in two ways. The symbol h refers to the cost of holding a single unit
of inventory for some fixed time, typically one year. It can also be thought of as an interest
rate, which is then multiplied by the unit cost of the item. I think the most appropriate value of
the interest rate is the rate of return you expect to earn in this business. I did a little research on
the web and it seems that for businesses of this type, a 15 percent annual return on investment
seems to be about right. We’ll use that to figure the holding cost.” Since Betty’s cost of each
bracelet was $30, this resulted in an annual holding cost of h = (.15)(30) = $4.50 annually.
Since the holding cost was based an annual interest rate, the sales rate had to be
yearly as well. This translated to an annual sales rate of (12)(75) = 900. Plugging these
numbers into the formula gave a lot size of
1 2 2 1 150 2 1 900 2
5 245. This made
Ä
4.5
more sense to Betty than ordering 75 bracelets every month. This translates to a little
more than a three month supply.
Once Betty got the idea, it didn’t take her long to set up a spreadsheet for all the
items in the store. For example, a similar analysis of the Lego sets resulted in an EOQ
value of 80 Lego sets. That seemed all well and good until her first order arrived.
nah77856_ch04.indd 239
17/12/14 6:39 PM
240 Chapter Four Inventory Control ­Subject to Known ­Demand
Uh oh, she thought, these things are pretty bulky. They filled up almost half of her
backroom storage area. She realized that there was a little more to figuring out the right
lot sizes than just applying the EOQ formula.
Betty asked Bob what she should do about the fact that the Lego order took up
so much space. “Well, I’m not sure,” answered Bob. “I guess you shouldn’t order so
many Lego sets.” “Well thanks for nothing, genius.” Betty responded. “I know that
now, but just how many should I order?”
So Bob did his research, and found out this was a tougher problem than figuring the
EOQ. As a first cut, Bob read that if the value of each item were proportional to the space
it consumed, the solution would be pretty simple. Realizing that this was only an approximation, he figured it would be better than nothing, and would hopefully satisfy Betty.
“Ok,” Bob said, “let’s consider your spreadsheet that computes the EOQ values for all
of the items in the store. Now let’s add a column to the spreadsheet that indicates the cubic
feet of space consumed by each item. Multiply the two for each item and add up the total.”
This was an easy calculation once Betty was able to approximate the space requirements for each of the items. Multiplying the two columns and adding resulted in a total
space requirement of 120,000 cubic feet. Betty had only 50,000 cubic feet of storage
space, so Bob suggested that each of the order quantities be reduced by multiplying
by the constant 50,000/120,000 = .4167. “If you reduce all of your order quantities by
around 60 percent, you shouldn’t run of the space,” Bob recommended.
“Bob, that’s great! That will be a big help,” said Betty. Betty added another column
to her spreadsheet that reduced all of the lot sizes by 60 percent. But as she started
looking at the numbers, something bothered her.
“You know, Bob, this will solve my space problem, but there’s something about it
that doesn’t make sense to me. Reducing the lot size of the Lego sets from 80 to 32
sounds about right, but why should I reduce the lot size for the bracelets? They hardly
take up any room at all.”
Jeez, she’s absolutely right on the money about that, thought Bob. You really wanted
to reduce the lot sizes of the bulky items a lot more than the small items. But how do
you do that? Bob consulted his friend Phil, who has a PhD in operations research. Phil
explained that this was a nontrivial optimization problem that involved finding the
value of something called the Lagrange multiplier. Given Betty’s spreadsheet, this was
a piece of cake for Phil. Once the calculation was completed, the lot sizes for the bulky
items were indeed reduced much more than those of the smaller items.
By applying a few basic concepts from inventory control, Betty was able to get a
handle on the problem of managing her stock. Her willingness to apply scientific principles meant that Betty was well on her way to creating a successful business.
4.11 Historical Notes and Additional Topics
The interest in using mathematical models to control the replenishment of inventories
dates back to the early part of the 20th century. Ford Harris (1915) is generally ­credited
with the development of the simple EOQ model. R. H. Wilson (1934) is also recognized
for his analysis of the model. The procedure we suggest for the all-units quantity discount
model appears to be due to Churchman, Ackoff, and Arnoff (1957), and the incremental
discount model appears to be due to Hadley and Whitin (1963). Kenneth Arrow provides
an excellent discussion of the economic motives for holding inventories in Chapter 1 of
Arrow, Karlin, and Scarf (1958).
Hadley and Whitin (1963) also seem to have been the first to consider budget- and
space-constrained problems, although several researchers have studied the problem
nah77856_ch04.indd 240
17/12/14 6:39 PM
4.12 Summary 241
subsequently. Rosenblatt (1981) derives results similar to those of Section 4.8 and
­con­siders several extensions not treated in this section.
Academic interest in inventory management problems took a sudden upturn in the
late 1950s and early 1960s with the publication of a number of texts in addition to those
just mentioned, including Whitin (1957); Magee and Boodman (1967); ­Bowman and
Fetter (1961); Fetter and Dalleck (1961); Hanssmann (1961); Star and Miller (1962);
Wagner (1962); and Scarf, Gilford, and Shelly (1963).
By and large, the vast majority of the published research on inventory systems since
the 1960s has been on stochastic models, which will be the subject of Chap­­­ter 5. Extensions of the EOQ model have been considered more ­recently by Barbosa and Friedman
(1978) and Schwarz and Schrage (1971), for example.
4.12 Summary
nah77856_ch04.indd 241
This chapter presented several popular models used to control inventories when the ­demand is
known. We discussed the various economic motives for holding inventories, which include economies of scale, uncertainties, speculation, and logistics. In addition, we mentioned some of the
physical characteristics of inventory systems that are i­mportant in determining the complexity
and applicability of the models to real problems. These include demand, lead time, ­review time,
back-order/lost-sales ­assumptions, and the changes that take place in the inventory over time.
There are three significant classes of costs in inventory management. These are holding or
carrying costs, order or setup costs, and penalty cost for not meeting d­ emand. The holding
cost is usually expressed as the product of an interest rate and the cost of the item.
The grandfather of all inventory control models is the simple EOQ model, in which
­demand is assumed to be constant, no stock-outs are permitted, and only h­ olding and order
costs are present. The optimal batch size is given by the classic square root formula. The
first extension of the EOQ model we considered was to the case in which items are produced
­internally at a finite production rate. We showed that the ­optimal batch size in this case could
be obtained from the EOQ formula with a ­modified holding cost.
Two types of quantity discounts were considered: all-units, in which the discounted price is valid
on all the units in an order, and incremental, in which the discount is ­applied only to the additional
units beyond the breakpoint. The optimal solution to the all-units case will often occur at a breakpoint, whereas the optimal solution to the ­incremental case will almost never occur at a breakpoint.
In most real systems, the inventory manager cannot ignore the interactions that exist
­between products. These interactions impose constraints on the system; these might arise
because of limitations in the space available to store inventory or in the budget available
to purchase items. We considered both cases, and showed that when the ratio of the item
value or space consumed by the item over the holding cost is the same for all items, a solution to the constrained problem can be obtained easily. When this condition is not met, the
formulation requires introducing a Lagrange multiplier. The correct value of the Lagrangian
­multiplier can be found by trial and error or by some type of search. The presence of the
­multiplier reduces lot sizes by effectively i­ncreasing holding costs.
The finite production rate model of Section 4.6 was extended to multiple products under the
assumption that all the products are produced on a single machine. Assuming a r­ otation cycle
policy in which one lot of each product is produced each cycle, we showed how to ­determine
the cycle time minimizing the sum of annual setup and holding costs for all products. Rotation
cycles provide a straightforward way to schedule production on one ­machine, but may be suboptimal when demand or production rates differ widely or setup costs are sequence dependent.
The chapter concluded with a brief overview of several commercial inventory control
systems. There are dozens of products available ranging in price from under $100 to tens of
thousands of dollars. Many of the products designed to run on personal computers are parts
of integrated accounting systems. Most of these products do not include item forecasting,
lot sizing, and reorder point calculations.
17/12/14 6:39 PM
nah77856_ch04.qxd
1/17/08
3:44 PM
Page 240
240 Chapter Four
Inventory Control
242 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
Additional
Problems on
Deterministic
Inventory
Models
31. Peet’s Coffees in Menlo Park, California, sells Melitta Number 101 coffee filters at a
fairly steady rate of about 60 boxes of filters monthly. The filters are ordered from a
supplier in Trenton, New Jersey. Peet’s manager is interested in applying some inventory
theory to determine the best replenishment strategy for the filters.
Peet’s pays $2.80 per box of filters and estimates that fixed costs of employee time
for placing and receiving orders amount to about $20. Peet’s uses a 22 percent annual
interest rate to compute holding costs.
a. How large a standing order should Peet’s have with its supplier in Trenton, and how
often should these orders be placed?
b. Suppose that it takes three weeks to receive a shipment. What inventory of filters
should be on hand when an order is placed?
c. What are the average annual holding and fixed costs associated with these filters,
assuming they adopt an optimal policy?
d. The Peet’s store in Menlo Park is rather small. In what way might this affect the
solution you recommended in part (a)?
32. A local supermarket sells a popular brand of shampoo at a fairly steady rate of 380 bottles per month. The cost of each bottle to the supermarket is 45 cents, and the cost of
placing an order has been estimated at $8.50. Assume that holding costs are based on
a 25 percent annual interest rate. Stock-outs of the shampoo are not allowed.
a. Determine the optimal lot size the supermarket should order and the time between
placements of orders for this product.
b. If the procurement lead time is two months, find the reorder point based on the onhand inventory.
c. If the item sells for 99 cents, what is the annual profit (exclusive of overhead and
labor costs) from this item?
33. Diskup produces a variety of personal computer products. High-density 3.5-inch disks
are produced at a rate of 1,800 per day and are shipped out at a rate of 800 per day. The
disks are produced in batches. Each disk costs the company 20 cents, and the holding
costs are based on an 18 percent annual interest rate. Shortages are not permitted. Each
production run of a disk type requires recalibration of the equipment. The company
estimates that this step costs $180.
a. Find the optimal size of each production run and the time between runs.
b. What fraction of the time is the company producing high-density 3.5-inch disks?
c. What is the maximum dollar investment in inventory that the company has in these disks?
34. Berry Computer is considering moving some of its operations overseas in order to reduce labor costs. In the United States, its main circuit board costs Berry $75 per unit
to produce, while overseas it costs only $65 to produce. Holding costs are based on a
20 percent annual interest rate, and the demand has been a fairly steady 200 units per
week. Assume that setup costs are $200 both locally and overseas. Production lead
times are one month locally and six months overseas.
a. Determine the average annual costs of production, holding, and setup at each location, assuming that an optimal solution is employed in each case. Based on these
results only, which location is preferable?
b. Determine the value of the pipeline inventory in each case. (The pipeline inventory is
the inventory on order.) Does comparison of the pipeline inventories alter the conclusion reached in part (a)?
nah77856_ch04.indd 242
17/12/14 6:39 PM
nah77856_ch04.qxd
1/17/08
3:44 PM
Page 241
Additional Problems on Deterministic Inventory Models 241
Additional Problems on Deterministic Inventory Models 243
c. Might considerations other than cost favor local over overseas production?
35. A large producer of household products purchases a glyceride used in one of its
deodorant soaps from outside of the company. It uses the glyceride at a fairly steady
rate of 40 pounds per month, and the company uses a 23 percent annual interest rate to
compute holding costs. The chemical can be purchased from two suppliers, A and B.
A offers the following all-units discount schedule:
Order Size
0 � Q � 500
500 � Q � 1,000
1,000 � Q
Price per Pound
$1.30
1.20
1.10
whereas B offers the following incremental discount schedule: $1.25 per pound for all
orders less than or equal to 700 pounds, and $1.05 per pound for all incremental
amounts over 700 pounds. Assume that the cost of order processing for each case is
$150. Which supplier should be used?
36. The president of Value Filters became very enthusiastic about using EOQs to plan the
sizes of her production runs, and instituted lot sizing based on EOQ values before she
could properly estimate costs. For one particular filter line, which had an annual demand of 1,800 units per year and which was valued at $2.40 per unit, she assumed a
holding cost based on a 30 percent annual interest rate and a setup cost of $100. Some
time later, after the cost accounting department had time to perform an analysis, it
found that the appropriate value of the interest rate was closer to 20 percent and the
setup cost was about $40. What was the additional average annual cost of holding and
setup incurred from the use of the wrong costs?
37. Consider the carload discount schedule pictured in Figure 4–13 (page 225). Suppose
M � 500 units, C � $10 per unit, and a full carload of 500 units costs $3,000.
a. Develop a graph of the average cost per unit, C(Q)Q, assuming this schedule.
b. Suppose that the units are consumed at a rate of 800 per week, order setup cost is
$2,500, and holding costs are based on an annual interest charge of 22 percent.
Graph the function G(Q) � C(Q)Q � KQ � I(C(Q)Q)Q2 and find the
optimal value of Q. (Assume that 1 year � 50 weeks.)
c. Repeat part (b) for � 1,000 per week and K � $1,500.
38. Harold Gwynne is considering starting a sandwich-making business from his dormitory room in order to earn some extra income. However, he has only a limited budget
of $100 to make his initial purchases. Harold divides his needs into three areas: breads,
meats and cheeses, and condiments. He estimates that he will be able to use all the
products he purchases before they spoil, so perishability is not a relevant issue. The demand and cost parameters are as follows.
Weekly demand
Cost per unit
Fixed order cost
nah77856_ch04.indd 243
Breads
Meats
and
Cheeses
Condiments
6 packages
$0.85
$12
12 packages
$3.50
$8
2 pounds
$1.25
$10
17/12/14 6:39 PM
nah77856_ch04.qxd
1/17/08
3:44 PM
Page 242
244 Chapter
­Subject to Known Demand
­Demand
242 Chapter Four Inventory
Four Inventory Control Subject
The choice of these fixed costs is based on the fact that these items are purchased at
different locations in town. They include the cost of Harold’s time in making the purchase. Assume that holding costs are based on an annual interest rate of 25 percent.
a. Find the optimal quantities that Harold should purchase of each type of product so
that he does not exceed his budget.
b. If Harold could purchase all the items at the same location, would that alter your solution? Why?
39. Mike’s Garage, a local automotive service and repair shop, uses oil filters at a fairly
steady rate of 2,400 per year. Mike estimates that the cost of his time to make and
process an order is about $50. It takes one month for the supplier to deliver the oil filters to the garage, and each one costs Mike $5. Mike uses an annual interest rate of
25 percent to compute his holding cost.
a. Determine the optimal number of oil filters that Mike should purchase, and the optimal time between placement of orders.
b. Determine the level of on-hand inventory at the time a reorder should be placed.
c. Assuming that Mike uses an optimal inventory control policy for oil filters, what is
the annual cost of holding and order setup for this item?
40. An import/export firm has leased 20,000 cubic feet of storage space to store six items
that it imports from the Far East. The relevant data for the six items it plans to store in
the warehouse are
Item
Annual Demand
(units)
DVD
800
32-inch flat screen TV 1,600
Blank DVD’s
8,000
(box of 10)
Blank CD’s
12,000
(box of 50)
Compact stereo
400
Telephone
1,200
Cost per Unit
Space per Unit
(feet 3)
$200.00
150.00
30.00
12
18
3
18.00
2
250.00
12.50
24
3
Setup costs for each product amount to $2,000 per order, and holding costs are
based on a 25 percent annual interest rate. Find the optimal order quantities for these
six items so that the storage space is never exceeded. (Hint: Use a cell location for the
Lagrange multiplier and experiment with different values until the storage constraint
is satisfied as closely as possible.)
41. A manufacturer of greeting cards must determine the size of production runs for a certain popular line of cards. The demand for these cards has been a fairly steady 2 million per year, and the manufacturer is currently producing the cards in batch sizes of
50,000. The cost of setting up for each production run is $400.
Assume that for each card the material cost is 35 cents, the labor cost is 15 cents, and
the distribution cost is 5 cents. The accounting department of the firm has established
an interest rate to represent the opportunity cost of alternative investment and storage
costs at 20 percent of the value of each card.
a. What is the optimal value of the EOQ for this line of greeting cards?
b. Determine the additional annual cost resulting from using the wrong production lot size.
nah77856_ch04.indd 244
17/12/14 6:39 PM
nah77856_ch04.qxd
1/17/08
3:44 PM
Page 243
2.1 Level
One
Head 245
AdditionalProblems
Problemson
onDeterministic
Deterministic
Inventory
Models 245
Additional
Inventory
Models
243
42. Suppose that in Problem 41 the firm decides to account for the fact that the production
rate of the cards is not infinite. Determine the optimal size of each production run
assuming that cards are produced at the rate of 75,000 per week.
43. Pies ‘R’ Us bakes its own pies on the premises in a large oven that holds 100 pies. They
sell the pies at a fairly steady rate of 86 per month. The pies cost $2 each to make. Prior
to each baking, the oven must be cleaned out, which requires one hour’s time for four
workers, each of whom is paid $8 per hour. Inventory costs are based on an 18 percent
annual interest rate. The pies have a shelf life of three months.
a. How many pies should be baked for each production run? What is the annual cost
of setup and holding for the pies?
b. The owner of Pies ‘R’ Us is thinking about buying a new oven that requires one-half
the cleaning time of the old oven and has a capacity twice as large as the old one.
What is the optimal number of pies to be baked each time in the new oven?
c. The net cost of the new oven (after trading in the old oven) is $350. How many years
would it take for the new oven to pay for itself?
44. The Kirei-Hana Japanese Steak House in San Francisco consumes 3,000 pounds of
sirloin steak each month. Yama Hirai, the new restaurant manager, recently completed
an MBA degree. He learned that the steak was replenished using an EOQ value of
2,000 pounds. The EOQ value was computed assuming an interest rate of 36 percent
per year. Assume that the current cost of the sirloin steak to the steak house is $4 per
pound.
a. What is the setup cost used in determining the EOQ value?
Mr. Hirai received an offer from a meat wholesaler in which a discount of 5 percent
would be given if the steak house purchased the steak in quantities of 3,000 pounds or
more.
b. Should Mr. Hirai accept the offer from the wholesaler? If so, how much can be
saved?
c. Because of Mr. Hirai’s language problems, he apparently misunderstood the offer.
In fact, the 5 percent discount is applied only to the amounts ordered over 3,000
pounds. Should this offer be accepted, and if so, how much is now saved?
45. Green’s Buttons of Rolla, Missouri, supplies all the New Jersey Fabrics stores with eight
different styles of buttons for men’s dress shirts. The plastic injection molding machine
can produce only one button style at a time and substantial time is required to reconfigure the machine for different button styles. As Green’s has contracted to supply fixed
quantities of buttons for the next three years, its demand can be treated as fixed and
known. The relevant data for this problem appear in the following table.
nah77856_ch04.indd 245
Button
Type
Annual
Sales
Production Rate
(units/day)
Setup Time
(hours)
Variable
Cost
A
B
C
D
E
F
G
H
25,900
42,000
14,400
46,000
12,500
75,000
30,000
18,900
4,500
5,500
3,300
3,200
1,800
3,900
2,900
1,200
6
4
8
4
3
6
1
3
$0.003
0.002
0.008
0.002
0.010
0.005
0.004
0.007
17/12/14 6:39 PM
nah77856_ch04.qxd
1/17/08
3:44 PM
Page 244
244 Chapter Four Inventory Control Subject to Known Demand
244 Chapter Four
Inventory Control
246 Chapter
Four Inventory
Control Subject
­Subject to
to Known
Known Demand
­Demand
Appendix 4–A 246
Assume 250 working days per year. Green’s accounting department has established
an 18
percent
annual
interest
theGreen’s
cost ofaccounting
capital and department
a 3 percent has
annual
interest
Assume
250
working
daysrate
per for
year.
established
rate
to
account
for
storage
space.
Setup
costs
are
$20
per
hour
required
to
reconfigure
an 18 percent annual interest rate for the cost of capital and a 3 percent annual interest
the
a new style.
that are
the $20
firmper
decides
to use a to
rotation
cycle
rate equipment
to account for
for storage
space.Suppose
Setup costs
hour required
reconfigure
policy
for
production
of
the
buttons.
the equipment for a new style. Suppose that the firm decides to use a rotation cycle
policy
forisproduction
the buttons.
a.
What
the optimalofrotation
cycle time?
b.
therotation
lots be?cycle time?
a. How
Whatlarge
is theshould
optimal
c.
What
is
the
average
annual
cost of holding and setups at the optimal solution?
b. How large should the lots be?
d.
obligations
might
Green’s have
with New
Jersey
Fabrics
that would
c. What
What contractual
is the average
annual cost
of holding
and setups
at the
optimal
solution?
prevent
them
from
implementing
the
policy
you
determined
in
parts
(a)
and
(b)?
d. What contractual obligations might Green’s have with New Jersey Fabrics that
would
More
specifically,
if
Green’s
agreed
to
make
three
shipments
per
year
for
each
button
prevent them from implementing the policy you determined in parts (a) and (b)?
style,
production
policyagreed
wouldtoyou
recommend?
More what
specifically,
if Green’s
make
three shipments per year for each button
style, what production policy would you recommend?
Appendix 4–A
Appendix 4–A
Mathematical Derivations for Multiproduct
Mathematical
Derivations
Constrained EOQ
Systemsfor Multiproduct
Constrained
Systems
Consider the standardEOQ
multiproduct
EOQ system with a budget constraint, which was
discussed
in Section
Mathematically,
problem
find values
of thewhich
variables
Consider the
standard4.8.
multiproduct
EOQ the
system
with isa to
budget
constraint,
was
Q
,
Q
,
.
.
.
,
Q
to
1
2
n
discussed in Section 4.8. Mathematically, the problem is to find values of the variables
Q1, Q2, . . . , Qn to
subject to
n hQ
Kii
i i
Minimize a c
d
n
i1 hi2Qi
KQii i
Minimize a c
d
Qi
i1 2
c Q … C.
a
n i i
i1
a ciQi … C.
i1
Let EOQi be the respective unconstrained
EOQ values. Then there are two
possibilities:
Let EOQi be the respective unconstrained EOQ values. Then there are two
possibilities:
n
c EOQi … C,
(1)
a
n i
i1
(1)
a ciEOQi … C,
subject to
n
c EOQi
a
n i
i1
a ciEOQi
i1 is the
solution
n
i1
7 C.
(2)
7 C.
(2)
If Equation (1) holds, the optimal
trivial solution; namely, set Qi EOQi. If
Equation
(2)(1)holds,
we are guaranteed
thattrivial
the constraint
is binding
optimal
If Equation
holds,then
the optimal
solution is the
solution; namely,
set at
Qi the
EOQ
i. If
solution.
That
means
that
the
constraint
may
be
written
Equation (2) holds, then we are guaranteed that the constraint is binding at the optimal
solution. That means that the constraint nmay be written
c Q C.
a
n i i
i1
a ciQi C.
i1
nah77856_ch04.indd 246
17/12/14 6:39 PM
nah77856_ch04.qxd
1/17/08
3:44 PM
Page 245
Appendix 4–A 247
4–A 245
Appendix
In this case, we introduce the Lagrange multiplier , and the problem is now to find
Q1, Q2, . . . , Qn, and to solve the unconstrained problem:
n
n
hiQi Kii
Minimize G(Q1, Q2, . . . , Qn, u) a a
b u a (ciQi C).
2
Qi
i1
i1
Necessary conditions for optimality are that
0G
0
0Qi
for i 1, . . . , n
and
0G
0.
0u
The first n conditions give
hi Kii
2 uci 0 for i 1, . . . , n.
2
Qi
Rearranging terms, we get
Qi 2Kii
A hi 2uci
for i 1, . . . , n,
and we also have the final condition
a ciQi C.
n
i1
Now consider the case where cihi ch independent of i. By dividing the numerator
and the denominator by hi , we may write
Qi 2Kii
1
A hi A 1 2uc>h
EOQi
1
A 1 2uc>h
EOQim,
where
m
1
.
A 1 2uc>h
Substituting this expression for Qi into the constraint gives
a ciEOQim C
n
i1
or
m
C
a ciEOQi
n
.
i1
nah77856_ch04.indd 247
17/12/14 6:39 PM
nah77856_ch04.qxd
1/17/08
3:44 PM
Page 246
248 Chapter
Four Inventory
­Subject to Known Demand
­Demand
246 Chapter Four
Inventory Control Subject
Appendix 4–B
Glossary of Notation for Chapter 4
c Proportional order cost.
EOQ Economic order quantity (optimal lot size).
G(Q) Average annual cost associated with lot size Q.
h Holding cost per unit time.
h Modified holding cost for finite production rate model.
I Annual interest rate used to compute holding cost.
K Setup cost or fixed order cost.
Demand rate (units per unit time).
P Production rate for finite production rate model.
Q Lot size or size of the order.
si Setup time for product i (refer to Section 4.9).
T Cycle time; time between placement of successive orders.
Order lead time.
Lagrange multiplier for space-constrained model (refer to Section 4.8).
wi Space consumed by one unit of product i (refer to Section 4.8).
W Total space available (refer to Section 4.8).
Bibliography
Arrow, K. A.; S. Karlin; and H. Scarf, eds. Studies in the
Mathematical Theory of Inventory P
­ roduction. Stanford,
CA: Stanford University Press, 1958.
Barbosa, L. C., and M. Friedman. “Deterministic I­ nventory
Lot Size Models—A General Root Law.” Management
­Science 23 (1978), pp. 820–29.
Bowman, E. H., and R. B. Fetter. Analysis for Production
Management. New York: McGraw‑Hill/Irwin, 1961.
Bibliography
Churchman,
R. L. Ackoff;
and E.eds.
L. Arnoff.
Arrow,
K. A.;C.S.W.;
Karlin;
and H. Scarf,
StudiesIntroduction
in the
to
Operations
Research.
New
York:
John
Wiley
Sons, 1957.
Mathematical Theory of Inventory Production.&Stanford,
Donnelly,
H. “Technology
Awards:
Recognizing Retailing’s
CA:
Stanford
University Press,
1958.
Best.”
Stores
76,
no.
10
(1994),
pp.
52–56.
Barbosa, L. C., and M. Friedman. “Deterministic
Inventory
Fetter,
R.
B.,
and
W.
C.
Dalleck.
Decision
for Inventory
Lot Size Models—A General Root Law.”Models
Management
Management.
New York:
­McGraw-Hill/Irwin, 1961.
Science
23 (1978),
pp. 820–29.
Hadley,
G.
J.,
and
T.
M.
Whitin.
of for
Inventory
Bowman, E. H., and R. B. Fetter.Analysis
Analysis
Production
­Systems. Englewood
Cliffs,
NJ: Prentice Hall, 1963.
Management.
New York:
McGraw-Hill/Irwin,
1961.
Hanssmann, C.
F. “A
of Inventory
fromIntroducthe
Churchman,
W.;Survey
R. L. Ackoff;
and E.Theory
L. Arnoff.
­Operations
Research
Viewpoint.”
Progress
in ­Operations
tion
to Operations
Research.
New In
York:
John Wiley
& Sons,
­Research 1, ed. R. L. Ackoff. New York: John Wiley &
1957.
Sons, 1961.
Donnelly,
H. “Technology Awards: Recognizing Retailing’s
Harris,Stores
F. W. Operations
and Cost
Best.”
76, no. 10 (1994),
pp.(Factory
52–56. Management
­Series). Chicago: Shaw, 1915.
Fetter, R. B., and W. C. Dalleck. Decision Models for
Magee,
J. Management.
F., and D. M. Boodman.
Planning and
Inventory
New York: Production
McGraw-Hill/Irwin,
Inventory
Control.
2nd
ed.
New
York:
McGraw-Hill,
1967.
1961.
Maxwell, W. L. “The Scheduling of Economic Lot Sizes.”
Hadley, G. J., and T. M. Whitin. Analysis of Inventory
Naval Research Logistics Quarterly 11 (1964), pp. 89–124.
Systems. Englewood Cliffs, NJ: Prentice Hall, 1963.
nah77856_ch04.indd 248
Muckstadt, J. A., and R. O. Roundy. “Analysis of Multi-Stage
Production Systems.” Chapter 2 in Logistics of Production
and Inventory, eds. S. C. Graves, A. H. G. Rinnooy Kan, and
P. H. Zipkin. Volume 4 of Handbooks in Operations Research
and Management Science, Amsterdam: North-Holland, 1993.
Rosenblatt, M. J. “Multi-Item Inventory System with
­Budgetary Constraint: A Comparison b­ etween the
­Lagrangian and the Fixed Cycle Approach.” International
Journal of Production Research 19 (1981), pp. 331–39.
Hanssmann,
F. “98%-Effective
“A Survey of Inventory
TheoryLot-Sizing
from the
Roundy, R. O.
Integer-Ratio
Operations
ResearchMulti-Retailer
Viewpoint.” InSystems.”
Progress Management
in Operations
for One Warehouse
Research
ed. R. L.
New York: John Wiley &
­Science 311,(1985),
pp.Ackoff.
1416–30.
Sons, 1961.
Scarf, H. E.; D. M. Gilford; and M. W. Shelly. Multistage
Harris, F. W. Operations and Cost (Factory Management
­Inventory Models and T
­ echniques. ­Stanford, CA: Stanford
Series). Chicago: Shaw, 1915.
University Press, 1963.
Magee, J. F., and D. M. Boodman. Production Planning and
Schwarz, L. B., and L. Schrage. “Optimal and Systems Myopic
Inventory Control. 2nd ed. New York: McGraw-Hill, 1967.
Policies for Multiechelon P
­ roduction/Inventory ­Assembly
Maxwell,
L. “The Scheduling
Economic
Lot Sizes.”
­Systems.” W.
Management
Science 21of
(1971),
pp. 1285–94.
Naval Research Logistics Quarterly 11 (1964), pp. 89–124.
Starr, M. K., and D. W. Miller. Inventory Control: Theory
Muckstadt,
J. A.,
and R. O.Cliffs,
Roundy.
of Multiand Practice.
Englewood
NJ:“Analysis
Prentice Hall,
1962.
Stage Production Systems.” Chapter 2 in Logistics of ProWagner, H. M. Statistical Management of Inventory Systems.
duction and Inventory, eds. S. C. Graves, A. H. G. Rinnooy
New York: John Wiley & Sons, 1962.
Kan, and P. H. Zipkin. Volume 4 of Handbooks in OperaWhitin,
T. M. The
of Inventory
Management.
Rev.
tions
Research
andTheory
Management
Science,
Amsterdam:
ed.
Princeton,
NJ:
Princeton
University
Press,
1957.
North-Holland, 1993.
Wilson, R. H.
Scientific Routine
for Stock
Control.”
Rosenblatt,
M.“A
J. “Multi-Item
Inventory
System
with BudHarvard
Business
Review
13
(1934),
pp.
116–28.
getary Constraint: A Comparison between the Lagrangian
17/12/14 6:39 PM
Chapter Five
Inventory Control
­Subject to Uncertain
Demand
“Knowing what you’ve got, knowing what you need, knowing what
you don’t—that’s inventory control.”
—Frank Wheeler in Revolutionary Road
Chapter Overview
Purpose
To understand how one deals with uncertainty (randomness) in the demand when
computing replenishment policies for a single inventory item.
Key Points
1. What is uncertainty and when should it be assumed? Uncertainty means that
­demand is a random variable. A random variable is defined by its probability distribution, which is generally estimated from a past history of demands. In practice, it is
common to assume that demand follows a normal distribution. When demand is
assumed normal, one only needs to estimate the mean, m, and variance, s2. Clearly,
demand is uncertain to a greater or lesser extent in all real-world applications.
What value, then, does the analysis of Chapters 3 and 4 have, where demand was
­assumed known? Chapter 3 focused on systematic or predictable changes in the
demand pattern, such as peaks and valleys. Chapter 4 results for single items are useful if the variance of demand is low relative to the mean. In this chapter we consider
items whose primary variation is due to uncertainty rather than predictable causes.
If demand is described by a random variable, it is unclear what the optimization
criterion should be, since the cost function is a random variable as well. To handle
this, we assume that the objective is to minimize expected costs. The use of the
expectation operator is justified by the law of large numbers from probability,
since an inventory control problem invariably spans many planning periods. The
law of large numbers guarantees that the arithmetic average of the incurred costs
and the expected costs grow close as the number of planning periods gets large.
2. The newsvendor model. Consider a news vendor that decides each morning how
many papers to buy to sell during the day. Since daily demand is highly variable,
it is modeled with a random variable, D. Suppose that Q is the number of papers
he purchases. If Q is too large, he is left with unsold papers, and if Q is too small,
some demands go unfilled. If we let co be the unit overage cost, and cu be the
249
nah77856_ch05.indd 249
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 249
250 Chapter Five Inventory Control Subject to Uncertain Demand
Chapter Five Inventory Control Subject to Uncertain Demand 249
unit underage cost, then we show that the optimal number of papers he should
purchase at the start of a day, say Q*, satisfies:
F(Q*) � cu(cu � co)
where F(Q*) is the cumulative distribution function of D evaluated at Q* (which
is the same as the probability that demand is less than or equal to Q*).
3. Lot size–reorder point systems. The newsvendor model is appropriate for a problem
that essentially restarts from scratch every period. Yesterday’s newspaper has no
value in the market, save for the possible scrap value of the paper itself. However,
most inventory control situations that one encounters in the real world are not
like this. Unsold items continue to have value in the marketplace for many
periods. For these cases we use an approach that is essentially an extension of
the EOQ model of Chapter 4.
The lot size–reorder point system relies on the assumption that inventories are
reviewed continuously rather than periodically. That is, the state of the system is
known at all times. The system consists of two decision variables: Q and R. Q is
the order size and R is the reorder point. That is, when the inventory of stock on
hand reaches R, an order for Q units is placed. The model also allows for a
positive order lead time, . It is the demand over the lead time that is the key
uncertainty in the problem, since the lead time is the response time of the
system. Let D represent the demand over the lead time, and let F(t) be the
cumulative distribution function of D. Cost parameters include a fixed order
cost K, a unit penalty cost for unsatisfied demand p, and a per unit per unit
time holding cost h. Interpret as the average annual demand rate (that is,
the expected demand over a year). Then we show in this section that the
optimal values of Q and R satisfy the following two simultaneous nonlinear
equations:
Q�
2[K � pn(R)]
h
B
1 � F(R) � Qhp.
The solution to these equations requires a back-and-forth iterative solution
method. We provide details of the method only when the lead time demand
distribution is normal. Convergence generally occurs quickly. A quick and dirty
approximation is to set Q � EOQ and solve for R in the second equation. This
will give good results in most cases.
4. Service levels in (Q, R) systems. We assume two types of service: Type 1 service
is the probability of not stocking out in the lead time and is represented by the
symbol . Type 2 service is the proportion of demands that are filled from stock
(also known as the fill rate) and is represented by the symbol . Finding the
optimal (Q, R) subject to a Type 1 service objective is very easy. One merely finds
R from F(R) � and sets Q � EOQ. Unfortunately, what one generally means
by service is the Type 2 criterion, and finding (Q, R) in that case is more difficult.
For Type 2 service, we only consider the normal distribution. The solution
requires using standardized loss tables, L(z), which are supplied in the back
of the book. As with the cost model, setting Q � EOQ and solving for R will
usually give good results if one does not want to bother with an iterative
procedure.
nah77856_ch05.indd 250
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 250
250 Chapter Five Inventory Control Subject to Uncertain Demand
Chapter Five Inventory Control Subject to Uncertain Demand 251
In this chapter, we consider the link between inventory control and
forecasting, and how one typically updates estimates of the mean and standard
deviation of demand using exponential smoothing. The section concludes with a
discussion of lead time variability, and how that additional uncertainty is taken
into account.
5. Periodic review systems under uncertainty. The newsvendor model treats a product
that perishes quickly (after one period). However, periodic review models also
make sense when unsold product can be used in future periods. In this case the
form of the optimal policy is known as an (s, S) policy. Let u be the starting
inventory in any period. Then the (s, S) policy is
If u � s, order to S (that is, order S � u).
If u � s, don’t order.
Unfortunately, finding the optimal values of (s, S) each period is much more
difficult than finding the optimal (Q, R) policy, and is beyond the scope of this
book. We also briefly discuss service levels in periodic review systems.
6. Multiproduct systems. Virtually all inventory control problems occurring in the
operations planning context involve multiple products. One issue that arises in
multiproduct systems is determining the amount of effort one should expend
managing each item. Clearly, some items are more valuable to the business than
others. The ABC classification system is one means of ranking items. Items are
sequenced in decreasing order of annual dollar volume of sales or usage.
Ordering the items in this way, and graphing the cumulative dollar volume gives
an exponentially increasing curve known as a Pareto curve. Typically, 20 percent
of the items account for 80 percent of the annual dollar volume (A items), the
next 30 percent of the items typically account for the next 15 percent of the
dollar volume (B items), and the final 50 percent of the items account for the
final 5 percent of the dollar volume (C items). A items should receive the most
attention. Their inventory levels should be reviewed often, and they should carry
a high service level. B items do not need such close scrutiny, and C items are
typically ordered infrequently in large quantities.
7. Other issues. The discussion of stochastic inventory models in this chapter barely
reveals the tip of the iceberg in terms of the vast quantity of research done on
this topic. Two important areas of research are multi-echelon inventory systems,
and perishable inventory systems. A multi-echelon inventory system is one in
which items are stored at multiple locations linked by a network. Supply chains,
discussed in detail in Chapter 6, are such a system. Another important area of
research are items that change during storage, thus affecting their useful
lifetime. One class of such items are perishable items. Perishable items have a
fixed lifetime known in advance, and include food, pharmaceuticals and
photographic film. A related problem is managing items subject to
obsolescence. Obsolescence differs from perishability in that the useful lifetime
of an item subject to obsolescence cannot be predicted in advance.
Mathematical models for analyzing such problems are quite complex and well
beyond the scope of this book.
The management of uncertainty plays an important role in the success of any firm.
What are the sources of uncertainty that affect a firm? A partial list includes uncertainty in consumer preferences and trends in the market, uncertainty in the availability
nah77856_ch05.indd 251
17/12/14 7:10 PM
252 Chapter Five Inventory Control Subject to Uncertain Demand
and cost of labor and resources, uncertainty in vendor resupply times, uncertainty in
weather and its ramifications on operations logistics, uncertainty of financial variables
such as stock prices and interest rates, and uncertainty of demand for products and
­services.
Before the terrible tragedy of September 11, 2001, many of us would arrive at the
airport no more than 30 or 40 minutes before the scheduled departure time of our
flight. Now we might arrive two hours before the flight. Increased airport security
has not only ­increased the average time required, it has increased the uncertainty of
this time. To compensate for this increased uncertainty, we arrive far in advance of
our scheduled departure time to provide a larger buffer time. This same principle will
apply when managing inventories.
The uncertainty of demand and its effect on inventory management strategies are
the subjects of this chapter. When a quantity is uncertain, it means that we cannot
­predict its value exactly in advance. For example, a department store cannot exactly
predict the sales of a particular item on any given day. An airline cannot exactly predict the number of people that will choose to fly on any given flight. How, then, can
these firms choose the number of items to keep in inventory or the number of flights to
schedule on any given route?
Although exact sales of an item or numbers of seats booked on a plane cannot
be predicted in advance, one’s past experience can provide useful information for
planning. As shown in Section 5.1, previous observations of any random phenomenon can be used to estimate its probability distribution. By properly quantifying the
consequences of incorrect decisions, a well-thought-out mathematical model of the
­system being studied will result in intelligent strategies. When uncertainty is present,
the ­objective is almost always to minimize expected cost or maximize expected
profits.
Demand uncertainty plays a key role in many industries, but some are more susceptible to business cycles than others. The world economy saw one of the worst recessions in recent times in 2008. The stock market eventually dropped to about half of
its high in 2007 and unemployment levels soared. The retailing industry in particular
is very sensitive to the vicissitudes of consumer demand. Low cost providers such as
Costco and Walmart fared well, but high-end retailers such as Nordstrom and Bloomingdales suffered. Matching supply and demand becomes even more critical in times
of recession.
As some level of demand uncertainty seems to characterize almost all inventory
management problems in practice, one might question the value of the deterministic
inventory control models discussed in Chapter 4. There are two reasons for studying
deterministic models. One is that they provide a basis for understanding the fundamental trade-offs encountered in inventory management. Another is that they may be good
approximations depending on the degree of uncertainty in the demand.
To better understand the second point, let D be the demand for an item over a given
period of time. We express D as the sum of two parts, DDet and DRan. That is,
D 5 DDet 1 DRan,
where
DDet 5 Deterministic component of demand,
DRan 5 Random component of demand.
nah77856_ch05.indd 252
17/12/14 7:10 PM
nah77856_ch05.qxd
252
1/9/08
11:42 AM
Page 252
Chapter Five Inventory Control Subject to Uncertain Demand
Chapter Five Inventory Control Subject to Uncertain Demand 253
There are a number of circumstances under which it would be appropriate to treat D
as being deterministic even though DRan is not zero. Some of these are
1. When the variance of the random component, DRan, is small relative to the magnitude of D.
2. When the predictable variation is more important than the random variation.
3. When the problem structure is too complex to include an explicit representation of
randomness in the model.
An example of circumstance 2 occurs in the aggregate planning problem. Although
the forecast error of the aggregate demands over the planning horizon may not be zero,
we are more concerned with planning for the anticipated changes in the demand than
for the unanticipated changes. An example of circumstance 3 occurs in material
requirements planning (treated in detail in Chapter 7). The intricacies of the relationships among various component levels and end items make it difficult to incorporate
demand uncertainty into the analysis.
However, for many items, the random component of the demand is too significant to
ignore. As long as the expected demand per unit time is relatively constant and the
problem structure not too complex, explicit treatment of demand uncertainty is desirable. In this chapter we will examine several of the most important stochastic inventory
models and the key issues surrounding uncertainty.1
Overview of Models Treated in This Chapter
Inventory control models subject to uncertainty are basically of two types: (1) periodic
review and (2) continuous review. (Recall the discussion at the start of Chapter 4.
Periodic review means that the inventory level is known at discrete points in time only,
and continuous review means that the inventory level is known at all times.) Periodic
review models may be for one planning period or for multiple planning periods. For
one-period models, the objective is to properly balance the costs of overage (ordering
too much) and underage (ordering too little). Single-period models are useful in
several contexts: planning for initial shipment sizes for high-fashion items, ordering
policies for food products that perish quickly, or determining run sizes for items with
short useful lifetimes, such as newspapers. Because of this last application, the singleperiod stochastic inventory model has come to be known as the newsvendor model. The
newsvendor model will be the first one considered in this chapter.
From this writer’s experience, the vast majority of computer-based inventory
control systems on the market use some variant of the continuous review models
treated in the remainder of this chapter. They are, in a sense, extensions of the EOQ
model to incorporate uncertainty. Their popularity in practice is attributable to several
factors. First, the policies are easy to compute and easy to implement. Second, the
models accurately describe most systems in which there is ongoing replenishment of
inventory items under uncertainty. A detailed discussion of service levels is included as
well. Because estimating penalty costs is difficult in practice, service level approaches
are more frequently implemented than penalty cost approaches.
Multiperiod stochastic inventory models dominate the professional literature on inventory theory. There are enough results in this fascinating research area to compose a
1
nah77856_ch05.indd 253
For those unfamiliar with it, the word stochastic is merely a synonym for random.
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 253
254 Chapter Five Inventory Control Subject to Uncertain Demand
5.1
The Nature of Randomness 253
volume in its own right. However, our goal in this book is to concentrate on methodology
that has been applied in the real world. Although they provide insight, multiperiod
stochastic models are rarely implemented. In addition, the level of mathematical
sophistication they require is beyond that of this book. For these two reasons, multiperiod
inventory models subject to uncertainty are not considered here.
5.1
THE NATURE OF RANDOMNESS
In order to clarify what the terms randomness and uncertainty mean in the context of
inventory control, we begin with an example.
Example 5.1
On consecutive Sundays, Mac, the owner of a local newsstand, purchases a number of copies
of The Computer Journal, a popular weekly magazine. He pays 25 cents for each copy and sells
each for 75 cents. Copies he has not sold during the week can be returned to his supplier for
10 cents each. The supplier is able to salvage the paper for printing future issues. Mac has kept
careful records of the demand each week for the Journal. (This includes the number of copies
actually sold plus the number of customer requests that could not be satisfied.) The observed
demands during each of the last 52 weeks were
15
14
8
6
8
13
19
11
9
7
11
12
9
6
5
12
11
12
11
4
15
18
9
9
4
15
15
22
18
17
19
17
4
10
18
9
19
7
0
14
10
14
8
14
15
9
14
11
12
8
16
17
There is no discernible pattern to these data, so it is difficult to predict the demand for the
Journal in any given week. However, we can represent the demand experience of this item as
a frequency histogram, which gives the number of times each weekly demand occurrence was
observed during the year. The histogram for this demand pattern appears in Figure 5–1.
One uses the frequency histogram to estimate the probability that the number of copies of
the Journal sold in any week is a specific value. These probability estimates are obtained by
dividing the number of times that each demand occurrence was observed during the year
by 52. For example, the probability that demand is 10 is estimated to be 252 .0385, and the
FIGURE 5–1
7
Observed frequency
Frequency histogram
for a 52-week history
of sales of The
Computer Journal
at Mac’s
6
5
4
3
2
1
0
nah77856_ch05.indd 254
0
2
4
6
8
10
12
14
Number of sales
16
18
20
22
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 254
5.1 The Nature of Randomness 255
254 Chapter Five Inventory Control Subject to Uncertain Demand
probability that the demand is 15 is 552 � .0962. The collection of all the probabilities is
known as the empirical probability distribution. Cumulative probabilities also can be estimated
in a similar way. For example, the probability that there are nine or fewer copies of the Journal
sold in any week is (1 � 0 � 0 � 0 � 3 � 1 � 2 � 2 � 4 � 6) � 1952 � .3654.
Although empirical probabilities can be used in subsequent analysis, they
are inconvenient for a number of reasons. First, they require maintaining a record of
the demand history for every item. This can be costly and cumbersome. Second, the
distribution must be expressed (in this case) as 23 different probabilities. Other items
may have an even wider range of past values. Finally, it is more difficult to compute
optimal inventory policies with empirical distributions.
For these reasons, we generally approximate the demand history using a continuous distribution. The form of the distribution chosen depends upon the history of past
demand and its ease of use. By far the most popular distribution for inventory applications is the normal. One reason is the frequency with which it seems to accurately
model demand fluctuations. Another is its convenience. The normal model of demand
must be used with care, however, as it admits the possibility of negative values. When
using the normal distribution to describe a nonnegative phenomenon such as demand,
the likelihood of a negative observation should be sufficiently small (less than .01
should suffice for most applications) so as not to be a factor.
A normal distribution is determined by two parameters: the mean and the variance
2. These can be estimated from a history of demand by the sample mean D
and the
sample variance s2. Let D1, D2, . . . , Dn be n past observations of demand. Then
D
�
s2 �
1 n
a Di,
n i�1
n
1
D)2.
a (Di � n � 1 i�1
For the data pictured in Figure 5–1 we obtain
D � 11.73,
s � 4.74.
The normal density function, f(x), is given by the formula
f(x) �
1
1 x�
exp ��� ��
2
22
2
for �q � x � �q.
We substitute D as the estimator for and s as the estimator for .
The relative frequency histogram is the same as the frequency histogram pictured in
Figure 5–1, except that the y-axis entries are divided by 52. In Figure 5–2 we show the
normal density function that results from the substitution we made, superimposed on
the relative frequency histogram.
In practice, exponential smoothing is used to recursively update the estimates of the
mean and the standard deviation of demand. The standard deviation is estimated using
the mean absolute deviation (MAD). Both exponential smoothing and MAD are
discussed in detail in Chapter 2. Let Dt be the estimate of the mean after observing
demand Dt, and let MADt be the estimate of the MAD. Then
D
t � Dt � (1 � )D
t�1,
MADt � Dt � Dt�1 � (1 � )MADt�1,
nah77856_ch05.indd 255
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 255
256 Chapter Five Inventory Control Subject to Uncertain Demand
5.2
Optimization Criterion 255
FIGURE 5–2
Frequency histogram
and normal
approximation
0.12
0.11
0.10
Relative frequency
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
–6 –4 –2
0
2
4
6
8 10 12 14 16 18 20 22 24 26 28
Number of sales
where 0 1 is the smoothing constant. For normally distributed demand
1.25 * MAD.
A smoothing constant of .1 is generally used to ensure stability in the estimates
(see Chapter 2 for a more complete discussion of the issues surrounding the choice of
the smoothing constant).
5.2
OPTIMIZATION CRITERION
In general, optimization in production problems means finding a control rule that
achieves minimum cost. However, when demand is random, the cost incurred is itself
random, and it is no longer obvious what the optimization criterion should be. Virtually all stochastic optimization techniques applied to inventory control assume that the
goal is to minimize expected costs.
The motivation for using the expected value criterion is that inventory control problems are generally ongoing problems. Decisions are made repetitively. The law of large
numbers from probability theory says that the arithmetic average of many observations
of a random variable will converge to the expected value of that random variable. In the
context of the inventory problem, if we follow a control rule that minimizes expected
costs, then the arithmetic average of the actual costs incurred over many periods will
also be a minimum.
In certain circumstances, the expected value may not be the best optimization
criterion. When a product is acquired only once and not on an ongoing basis, it is not
clear that minimizing expected costs is appropriate. In such a case, maximizing the
nah77856_ch05.indd 256
17/12/14 7:10 PM
nah77856_ch05.qxd
256
1/9/08
11:42 AM
Page 256
5.2 Optimization Criterion 257
Chapter Five Inventory Control Subject to Uncertain Demand
probability of some event (such as satisfying a proportion of the demand) is generally
more suitable. However, because of the ongoing nature of most production problems,
the expected value criterion is used in virtually all stochastic inventory control
applications.
Problems for Sections 5.1 and 5.2
1. Suppose that Mac only kept track of the number of magazines sold. Would this
give an accurate representation of the demand for the Journal? Under what circumstances would the actual demand and the number sold be close, and under
what circumstances would they differ by a substantial amount?
2. What is the difference between deterministic and random variations in the pattern
of demands? Provide an example of a real problem in which predictable variation
would be important and an example in which random variation would be
important.
3. Oakdale Furniture Company uses a special type of woodworking glue in the
assembly of its furniture. During the past 36 weeks, the following amounts of
glue (in gallons) were used by the company:
25
17
18
38
42
29
26
46
22
31
19
21
21
50
24
46
40
39
29
43
46
19
34
31
35
31
33
39
51
34
24
36
30
21
32
30
a. Compute the mean and the standard deviation of this sample.
b. Consider the following class intervals for the number of gallons used each
week:
Less than 20.
20–27.
28–33.
34–37.
38–43.
More than 43.
Determine the proportion of data points that fall into each of these intervals.
Compare these proportions to the probabilities that a normal variate with the
mean and the standard deviation you computed in part (a) falls into each of
these intervals. Based on the comparison of the observed proportions and those
obtained from assuming a normal distribution, would you conclude that the
normal distribution provides an adequate fit of these data? (This procedure is
essentially the same as a chi-square goodness-of-fit test.)
c. Assume that the numbers of gallons of glue used each week are independent
random variables, having the normal distribution with mean and standard
deviation computed in part (a). What is the probability that the total number of
gallons used in six weeks does not exceed 200 gallons? (Hint: The mean of a
sum of random variables is the sum of the means and the variance of a sum of
independent random variables is the sum of the variances.)
nah77856_ch05.indd 257
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 257
258 Chapter Five Inventory Control Subject to Uncertain Demand
5.3
The Newsboy Model 257
4. In Problem 3, what other probability distributions might accurately describe
Oakdale’s weekly usage of glue?
5. Rather than keeping track of each demand observation, Betty Sucasas, a member
of the marketing staff with a large company that produces a line of switches, has kept
only grouped data. For switch C9660Q, used in small power supplies, she has
observed the following numbers of units of the switch shipped over the last year.
Units Shipped
0–2,000
2,001–5,000
5,001–9,000
9,001–12,000
12,001–18,000
18,001–20,000
Number of
Weeks
3
6
12
17
10
4
Based on these observations, estimate the mean and the standard deviation of the
weekly shipments. (Hint: This is known as grouped data. For the purposes of
your calculation, assume that all observations occur at the midpoint of each
interval.)
6. a. Consider the Oakdale Furniture Company described in Problem 3. Under what
circumstances might the major portion of the usage of the glue be predictable?
b. If the demand were predictable, would you want to use a probability law to
describe it? Under what circumstances might the use of a probability model of
demand be justified even if the demand could be predicted exactly?
5.3
THE NEWSVENDOR MODEL
Let us return to Example 5.1, in which Mac wishes to determine the number of copies
of The Computer Journal he should purchase each Sunday. A study of the historical
data showed that the demand during any week is a random variable that is approximately normally distributed, with mean 11.73 and standard deviation 4.74. Each copy
is purchased for 25 cents and sold for 75 cents, and he is paid 10 cents for each unsold
copy by his supplier. One obvious solution is that he should buy enough to meet the
mean demand, which is approximately 12 copies. There is something wrong with this
solution. Suppose Mac purchases a copy that he does not sell. His out-of-pocket
expense is 25 cents � 10 cents � 15 cents. Suppose, on the other hand, that he is
unable to meet the demand of a customer. In that case, he loses 75 cents � 25 cents �
50 cents profit. Hence, there is a significantly greater penalty for not having enough
than there is for having too much. If he only buys enough to satisfy mean demand, he
will stock-out with the same frequency that he has an oversupply. Our intuition tells us
that he should buy more than the mean, but how much more? This question is answered
in this section.
Notation
This problem is an example of the newsvendor model, in which a single product is to be
ordered at the beginning of a period and can be used only to satisfy demand during that
nah77856_ch05.indd 258
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 258
5.3 The Newsvendor Model 259
258 Chapter Five Inventory Control Subject to Uncertain Demand
period. Assume that all relevant costs can be determined on the basis of ending inventory. Define
co Cost per unit of positive inventory remaining at the end of the period
(known as the overage cost).
cu Cost per unit of unsatisfied demand. This can be thought of as a cost
per unit of negative ending inventory (known as the underage cost).
In the development of the model, we will assume that the demand D is a continuous
nonnegative random variable with density function f(x) and cumulative distribution
function F(x). [A brief review of probability theory is given in Appendix 5–A. In
particular, both F(x) and f(x) are defined there.]
The decision variable Q is the number of units to be purchased at the beginning of
the period. The goal of the analysis is to determine Q to minimize the expected costs
incurred at the end of the period.
Development of the Cost Function
A general outline for analyzing most stochastic inventory problems is the following:
1. Develop an expression for the cost incurred as a function of both the random
variable D and the decision variable Q.
2. Determine the expected value of this expression with respect to the density function
or probability function of demand.
3. Determine the value of Q that minimizes the expected cost function.
Define G(Q, D) as the total overage and underage cost incurred at the end of the period when Q units are ordered at the start of the period and D is the demand. If Q units
are purchased and D is the demand, Q D units are left at the end of the period as long
as Q D. If Q D, then Q D is negative and the number of units remaining on
hand at the end of the period is 0. Notice that
max Q D, 0 e
if Q Ú D,
if Q … D.
QD
0
In the same way, maxD Q, 0 represents the excess demand over the supply, or
the unsatisfied demand remaining at the end of the period. For any realization of the
random variable D, either one or the other of these terms will be zero.
Hence, it now follows that
G(Q, D) co max(0, Q D) cu max (0, D Q).
Next, we derive the expected cost function. Define
G(Q) E(G(Q, D)).
Using the rules outlined in Appendix 5–A for taking the expected value of a function
of a random variable, we obtain
c G(Q) co
o
nah77856_ch05.indd 259
q
0
Q
0
max(0, Q x) f(x) dx cu
(Q x)f(x) dx cu
q
Q
q
0
max(0, x Q)f(x) dx
(x Q) f(x) dx.
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 259
260 Chapter Five Inventory Control Subject to Uncertain Demand
5.3
The Newsboy Model 259
Determining the Optimal Policy
We would like to determine the value of Q that minimizes the expected cost G(Q). In
order to do so, it is necessary to obtain an accurate description of the function G(Q).
We have that
Q
q
dG(Q)
co 1f(x) dx cu (1)f(x) dx
0
dQ
Q
co F(Q) cu(1 F(Q)).
(This is a result of Leibniz’s rule, which indicates how one differentiates integrals.
Leibniz’s rule is stated in Appendix 5–A.)
It follows that
d 2G(Q)
dQ2
(co cu)f(Q) 0
for all Q 0.
Because the second derivative is nonnegative, the function G(Q) is said to be convex
(bowl shaped). We can obtain additional insight into the shape of G(Q) by further
analysis. Note that
dG(Q)
dQ
Q0
co F(0) cu(1 F(0))
cu 0
since F(0) 0.
Since the slope is negative at Q 0, G(Q) is decreasing at Q 0. The function G(Q)
is pictured in Figure 5–3.
It follows that the optimal solution, say Q*, occurs where the first derivative of G(Q)
equals zero. That is,
G(Q*) (co cu)F(Q*) cu 0.
FIGURE 5–3
(thousands)
G ( Q)
Expected cost function
for newsvendor model
13
12
11
10
9
8
7
6
5
4
3
2
1
0
– 10 – 6 – 2 0
100
200
Q*
nah77856_ch05.indd 260
300
400
Q
17/12/14 7:10 PM
nah77856_ch05.qxd
260
1/9/08
11:42 AM
Page 260
5.3 The Newsvendor Model 261
Chapter Five Inventory Control Subject to Uncertain Demand
Rearranging terms gives
F(Q*) � cu(co � cu).
We refer to the right-hand side of the last equation as the critical ratio. Because cu
and co are positive numbers, the critical ratio is strictly between zero and one. This
implies that for a continuous demand distribution this equation is always solvable.
As F(Q*) is defined as the probability that the demand does not exceed Q*, the critical ratio is the probability of satisfying all the demand during the period if Q* units are
purchased at the start of the period. It is important to understand that this is not the
same as the proportion of satisfied demands. When underage and overage costs are
equal, the critical ratio is exactly one-half. In that case Q* corresponds to the median of
the demand distribution. When the demand density is symmetric (such as the normal
density), the mean and the median are the same.
Example 5.1
(continued)
Consider the example of Mac’s newsstand. From past experience, we saw that the weekly demand for the Journal is approximately normally distributed with mean � 11.73 and standard
deviation � 4.74. Because Mac purchases the magazines for 25 cents and can salvage unsold copies for 10 cents, his overage cost is co � 25 � 10 � 15 cents. His underage cost is the
profit on each sale, so that cu � 75 � 25 � 50 cents. The critical ratio is cu(co � cu) �
0.500.65 � .77. Hence, he should purchase enough copies to satisfy all the weekly demand
with probability .77. The optimal Q* is the 77th percentile of the demand distribution (see
Figure 5–4).
Using either Table A–1 or Table A–4 at the back of the book, we obtain a standardized value
of z � 0.74. The optimal Q is
Q* � z � � (4.74)(0.74) � 11.73
� 15.24 15.
Hence, he should purchase 15 copies every week.
FIGURE 5–4
Determination of the
optimal order quantity
for the newsvendor
example
f (x)
Area = .77
11.73
nah77856_ch05.indd 261
Q*
x
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 261
262 Chapter Five Inventory Control Subject to Uncertain Demand
5.3
The Newsboy Model 261
Optimal Policy for Discrete Demand
Our derivation of the newsvendor formula was based on the assumption that the demand in
the period was described by a continuous probability distribution. We noted a number
of reasons for the desirability of working with continuous distributions. However, in
some cases, and particularly when the mean demand is small, it may not be possible to
obtain an accurate representation of the observed pattern of demand using a continuous
distribution. For example, the normal approximation of the 52-week history of the demand for the Journal pictured in Figure 5–2 may not be considered sufficiently accurate
for our purposes.
The procedure for finding the optimal solution to the newsvendor problem when the
demand is assumed to be discrete is a natural generalization of the continuous case.
In the continuous case, the optimal solution is the value of Q that makes the distribution
function equal to the critical ratio cu(cu � co). In the discrete case, the distribution
function increases by jumps; it is unlikely that any of its values exactly equal the critical
ratio. The critical ratio will generally fall between two values of F(Q). The optimal
solution procedure is to locate the critical ratio between two values of F(Q) and choose
the Q corresponding to the higher value. (The fact that you always round Q up rather
than simply round it off can be easily proven mathematically. This is different from assuming that units are ordered in discrete amounts but demand is continuous, in which
case Q is rounded to the closest integer, as we did earlier.)
Example 5.2
We will solve the problem faced by Mac’s newsstand using the empirical distribution derived
from a one-year history of the demand, rather than the normal approximation of that demand
history. The empirical probabilities are obtained from Figure 5–1 by dividing each of the heights
by 52. We obtain
Q
f(Q)
F(Q)
Q
f(Q)
F(Q)
0
1
2
3
4
5
6
7
8
9
10
11
152
0
0
0
352
152
252
252
452
652
252
552
152 (.0192)
152 (.0192)
152 (.0192)
152 (.0192)
452 (.0769)
552 (.0962)
752 (.1346)
952 (.1731)
1352 (.2500)
1952 (.3654)
2152 (.4038)
2652 (.5000)
12
13
14
15
16
17
18
19
20
21
22
452
152
552
552
152
352
352
352
0
0
152
3052 (.5769)
3152 (.5962)
3652 (.6923)
4152 (.7885)
4252 (.8077)
4552 (.8654)
4852 (.9231)
5152 (.9808)
5152 (.9808)
5152 (.9808)
5252 (1.0000)
The critical ratio for this problem was .77, which corresponds to a value of F(Q) between
Q � 14 and Q � 15. Because we round up, the optimal solution is Q* � 15. Notice that this
is exactly the same order quantity obtained using the normal approximation.
Extension to Include Starting Inventory
In the derivation of the newsvendor model, we assumed that the starting inventory at the beginning of the period was zero. Suppose now that the starting inventory is some value u
and u � 0. The optimal policy for this case is a simple modification of the case u � 0.
Note that this extension would not apply to newspapers, but would be appropriate for a
product with a shelf life that exceeds one period.
nah77856_ch05.indd 262
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 262
Snapshot Application
USING INVENTORY MODELS TO MANAGE
THE SEED-CORN SUPPLY CHAIN AT SYNGENTA
Each year farmers plant tens of thousands of acres of
corn worldwide to meet the demands of a growing
population. Companies such as Minnesota-based Syngenta Seeds supply seed to these farmers. In the United
States alone, the market for corn seed is approximately
$2.3 billion annually. Syngenta is one of eight firms that
accounts for 73 percent of this market. Where do
Syngenta and its competitors obtain this seed? The answer is that it is produced by growing corn and harvesting the seeds. Corn grown for the purpose of harvesting
seed is known as seed-corn.
The problem of determining how much seed-corn to
plant is complicated by several factors. One is that there
are hundreds of different seed hybrids. Some hybrids do
better in warmer, more humid climates, while others do
better in cooler, dryer climates. The color, texture, sugar
content, and so forth, of the corn produced by different
hybrids varies as well. Farmers will not reuse a hybrid
that yielded disappointing results. Hence, annual
demand is hard to predict. In addition to facing uncertain demand, seed-corn producers also face uncertain
yields. Their seed-corn plantings are subject to the same
set of risks faced by all farmers: frost, draught, and heat
spells.
Syngenta must decide each season how much seedcorn to plant. Since demand for the seeds is uncertain,
the problem sounds like a straightforward application
of the newsvendor model with uncertain demand and uncertain supply. However, Syngenta’s decision problem
has an additional feature that makes it more complicated than an ordinary newsvendor problem. Syngenta,
along with many of its competitors, plants seed-corn in
both northern and southern hemispheres. Since the
hemispheric seasons run counter to each other, the
plantings are done at different times of the year. In particular, the seed-corn is planted in the spring season in
each hemisphere, so that the South American planting
occurs about six months after the North American
planting. This gives the company a second chance to
increase production levels in South America to make up
for shortfalls in North America, or decrease production
levels in South America when there are surpluses in
North America.
The problem of planning the size of the seed-corn
planting was tackled by a team of researchers from the
University of Iowa in collaboration with a vice president
in charge of supply at Syngenta.1 Using discrete approximations of the demand and yield distributions, they
were able to formulate the planning problem as a linear program, so that it could be solved on a firmwide
scale. A retrospective analysis showed that the company
could have saved upwards of $5 million using the
model. Also, the analysts were able to identify a systematic bias in the forecasts for seed generated by the firm
that resulted in consistent overproduction. The mathematical model is now used to help guide the firm on its
planting decisions each year.
1
Jones, P. C.; Kegler, G.; Lowe, T. J.; and Traub, R. D.
“Managing the Seed-Corn Supply Chain at Syngenta,”
Interfaces 33 (1), January-February 2003, pp. 80–90.
Consider the expected cost function G(Q), pictured in Figure 5–3. If u � 0, it just
means that we are starting at some point other than 0. We still want to be at Q* after
ordering, as that is still the lowest point on the cost curve. If u � Q*, this is accomplished by ordering Q* � u. If u � Q*, we are past where we want to be on the curve.
Ordering any additional inventory merely moves us up the cost curve, which results in
higher costs. In this case it is optimal to simply not order.
Hence the optimal policy when there is a starting inventory of u � 0 is
Order Q* � u if u � Q*.
Do not order if u � Q*.
Note that Q* should be interpreted as the order-up-to point rather than the order
quantity when u � 0. It is also known as a target or base stock level.
Example 5.2
(continued)
nah77856_ch05.indd 263
Let us suppose that in Example 5.2, Mac has received 6 copies of the Journal at the beginning
of the week from another supplier. The optimal policy still calls for having 15 copies on hand
after ordering, so now he would order the difference 15 � 6 � 9 copies. (Set Q* � 15 and
u � 6 to get the order quantity of Q* � u � 9.)
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 263
264 Chapter Five Inventory Control Subject to Uncertain Demand
5.3
The Newsboy Model 263
Extension to Multiple Planning Periods
The underlying assumption made in the derivation of the newsvendor model was that the
item “perished” quickly and could not be used to satisfy demand in subsequent
periods. In most industrial and retail environments, however, products are durable and
inventory left at the end of a period can be stored and used to satisfy future demand.
This means that the ending inventory in any period becomes the starting inventory
in the next period. Previously, we indicated how the optimal policy is modified when
starting inventory is present. However, when the number of periods remaining exceeds
one, the value of Q* also must be modified. In particular, the interpretation of both co
and cu will be different. We consider only the case in which there are infinitely many
periods remaining. The optimal value of the order-up-to point when a finite number of
periods remain will fall between the one-period and the infinite-period solutions.
In our derivation and subsequent analysis of the EOQ formula in Chapter 4, we saw
that the variable order cost c only entered into the optimization to determine the holding cost (h � Ic). In addition, we saw that all feasible operating policies incurred the
same average annual cost of replenishment, c. It turns out that essentially the same
thing applies in the infinite horizon newsvendor problem. As long as excess demand is
back-ordered, all feasible policies will just order the demand over any long period of
time. Similarly, as long as excess demand is back-ordered, the number of units sold
will just be equal to the demand over any long period of time. Hence, both cu and co
will be independent of both the proportional order cost c and the selling price of the
item. Interpret cu as the loss-of-goodwill cost and co as the holding cost in this case.
That this is the correct interpretation of the underage and overage costs is established
rigorously in Appendix 5–B.
Example 5.3
Let us return to Mac’s newsstand, described in Examples 5.1 and 5.2. Suppose that Mac is considering how to replenish the inventory of a very popular paperback thesaurus that is ordered
monthly. Copies of the thesaurus unsold at the end of a month are still kept on the shelves for
future sales. Assume that customers who request copies of the thesaurus when they are out of
stock will wait until the following month. Mac buys the thesaurus for $1.15 and sells it for
$2.75. Mac estimates a loss-of-goodwill cost of 50 cents each time a demand for a thesaurus
must be back-ordered. Monthly demand for the book is fairly closely approximated by a normal distribution with mean 18 and standard deviation 6. Mac uses a 20 percent annual interest rate to determine his holding cost. How many copies of the thesaurus should he purchase
at the beginning of each month?
Solution
The overage cost in this case is just the cost of holding, which is (1.15)(0.20)12 � 0.0192.
The underage cost is just the loss-of-goodwill cost, which is assumed to be 50 cents.
Hence, the critical ratio is 0.5(0.5 � 0.0192) � .9630. From Table A–1 at the back of this
book, this corresponds to a z value of 1.79. The optimal value of the order-up-to point
Q* � z � � (6)(1.79) � 18 � 28.74 29.
Example 5.3
(continued)
Assume that a local bookstore also stocks the thesaurus and that customers will purchase
the thesaurus there if Mac is out of stock. In this case excess demands are lost rather than
back-ordered. The order-up-to point will be different from that obtained assuming full
back-ordering of demand. In Appendix 5–B we show that in the lost sales case the underage
cost should be interpreted as the loss-of-goodwill cost plus the lost profit. The overage
cost should still be interpreted as the holding cost only. Hence, the lost sales solution for
this example gives cu � 0.5 � 1.6 � 2.1. The critical ratio is 2.1(2.1 � 0.0192) � .9909,
giving a z value of 2.36. The optimal value of Q in the lost sales case is Q* � z � �
(6)(2.36) � 18 � 32.16 32.
nah77856_ch05.indd 264
17/12/14 7:10 PM
nah77856_ch05.qxd
264
1/9/08
11:42 AM
Page 264
5.3 The Newsvendor Model 265
Chapter Five Inventory Control Subject to Uncertain Demand
Although the multiperiod solution appears to be sufficiently general to cover many
types of real problems, it suffers from one serious limitation: there is no fixed cost of
ordering. This means that the optimal policy, which is to order up to Q*, requires that
ordering take place in every period. In most real systems, however, there are fixed costs
associated with ordering, and it is not optimal to place orders each period. Unfortunately,
if we include a fixed charge for placing an order, it becomes extremely difficult to determine optimal operating policies. For this reason, we approach the problem of random
demand when a fixed charge for ordering is present in a different way. We will assume
that inventory levels are reviewed continuously and develop a generalization of the EOQ
analysis presented in Chapter 4. This analysis is presented in Section 5.4.
Problems for Section 5.3
7. A newsvendor keeps careful records of the number of papers he sells each day and the
various costs that are relevant to his decision regarding the optimal number of newspapers to purchase. For what reason might his results be inaccurate? What would he
need to do in order to accurately measure the daily demand for newspapers?
8. Billy’s Bakery bakes fresh bagels each morning. The daily demand for bagels is
a random variable with a distribution estimated from prior experience given by
Number of Bagels
Sold in One Day
0
5
10
15
20
25
30
35
Probability
.05
.10
.10
.20
.25
.15
.10
.05
The bagels cost Billy’s 8 cents to make, and they are sold for 35 cents each.
Bagels unsold at the end of the day are purchased by a nearby charity soup kitchen
for 3 cents each.
a. Based on the given discrete distribution, how many bagels should Billy’s bake
at the start of each day? (Your answer should be a multiple of 5.)
b. If you were to approximate the discrete distribution with a normal distribution,
would you expect the resulting solution to be close to the answer that you obtained
in part (a)? Why or why not?
c. Determine the optimal number of bagels to bake each day using a normal approximation. (Hint: You must compute the mean and the variance 2 of the
demand from the given discrete distribution.)
9. The Crestview Printing Company prints a particularly popular Christmas card once
a year and distributes the cards to stationery and gift shops throughout the United
States. It costs Crestview 50 cents to print each card, and the company receives
65 cents for each card sold.
Because the cards have the current year printed on them, those cards that are not
sold are generally discarded. Based on past experience and forecasts of current
nah77856_ch05.indd 265
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 265
5.3
266 Chapter Five Inventory Control Subject to Uncertain Demand
The Newsboy Model 265
buying patterns, the probability distribution of the number of cards to be sold
nationwide for the next Christmas season is estimated to be
Quantity Sold
100,000–150,000
150,001–200,000
200,001–250,000
250,001–300,000
300,001–350,000
350,001–400,000
400,001–450,000
Probability
.10
.15
.25
.20
.15
.10
.05
Determine the number of cards that Crestview should print this year.
10. Happy Henry’s car dealer sells an imported car called the EX123. Once every
three months, a shipment of the cars is made to Happy Henry’s. Emergency
shipments can be made between these three-month intervals to resupply the cars
when inventory falls short of demand. The emergency shipments require two
weeks, and buyers are willing to wait this long for the cars, but will generally go
elsewhere before the next three-month shipment is due.
From experience, it appears that the demand for the EX123 over a three-month
interval is normally distributed with a mean of 60 and a variance of 36. The cost of
holding an EX123 for one year is $500. Emergency shipments cost $250 per car
over and above normal shipping costs.
a. How many cars should Happy Henry’s be purchasing every three months?
b. Repeat the calculations, assuming that excess demands are back-ordered from
one three-month period to the next. Assume a loss-of-goodwill cost of $100 for
customers having to wait until the next three-month period and a cost of $50 per
customer for bookkeeping expenses.
c. Repeat the calculations, assuming that when Happy Henry’s is out of stock of
EX123s, the customer will purchase the car elsewhere. In this case, assume that
the cars cost Henry an average of $10,000 and sell for an average of $13,500.
Ignore loss-of-goodwill costs for this calculation.
11. Irwin’s sells a particular model of fan, with most of the sales being made in the
summer months. Irwin’s makes a one-time purchase of the fans prior to each
summer season at a cost of $40 each and sells each fan for $60. Any fans unsold
at the end of the summer season are marked down to $29 and sold in a special fall
sale. Virtually all marked-down fans are sold. The following is the number of sales
of fans during the past 10 summers: 30, 50, 30, 60, 10, 40, 30, 30, 20, 40.
a. Estimate the mean and the variance of the demand for fans each summer.
b. Assume that the demand for fans each summer follows a normal distribution,
with mean and variance given by what you obtained in part (a). Determine the
optimal number of fans for Irwin’s to buy prior to each summer season.
c. Based on the observed 10 values of the prior demand, construct an empirical
probability distribution of summer demand and determine the optimal number
of fans for Irwin’s to buy based on the empirical distribution.
d. Based on your results for parts (b) and (c), would you say that the normal
distribution provides an adequate approximation?
nah77856_ch05.indd 266
17/12/14 7:10 PM
nah77856_ch05.qxd
266
1/9/08
11:42 AM
Page 266
Chapter Five Inventory Control Subject to Uncertain Demand
5.4 Lot Size–Reorder Point Systems 267
12. The buyer for Needless Markup, a famous “high end” department store, must
decide on the quantity of a high-priced woman’s handbag to procure in Italy for the
following Christmas season. The unit cost of the handbag to the store is $28.50 and
the handbag will sell for $150.00. Any handbags not sold by the end of the season
are purchased by a discount firm for $20.00. In addition, the store accountants
estimate that there is a cost of $0.40 for each dollar tied up in inventory, as this
dollar invested elsewhere could have yielded a gross profit. Assume that this cost
is attached to unsold bags only.
a. Suppose that the sales of the bags are equally likely to be anywhere from 50
to 250 handbags during the season. Based on this, how many bags should the
buyer purchase? (Hint: This means that the correct distribution of demand is
uniform. You may solve this problem assuming either a discrete or a continuous
uniform distribution.)
b. A detailed analysis of past data shows that the number of bags sold is better
described by a normal distribution, with mean 150 and standard deviation 20.
Now what is the optimal number of bags to be purchased?
c. The expected demand was the same in parts (a) and (b), but the optimal order
quantities should have been different. What accounted for this difference?
5.4
LOT SIZE–REORDER POINT SYSTEMS
The form of the optimal solution for the simple EOQ model with a positive lead time
analyzed in Chapter 4 is: When the level of on-hand inventory hits R, place an order for
Q units. In that model the only independent decision variable was Q, the order quantity.
The value of R was determined from Q, , and . In what follows, we also assume that the
operating policy is of the (Q, R) form. However, when generalizing the EOQ analysis to
allow for random demand, we treat Q and R as independent decision variables.
The multiperiod newsvendor model was unrealistic for two reasons: it did not include
a setup cost for placing an order and it did not allow for a positive lead time. In most
real systems, however, both a setup cost and a lead time are present. For these reasons,
the kinds of models discussed in this section are used much more often in practice and,
in fact, form the basis for the policies used in many commercial inventory systems.
Note that Q in this section is the amount ordered, whereas Q in Section 5.3 was the
order-up-to point.
We make the following assumptions:
1. The system is continuous review. That is, demands are recorded as they occur, and
the level of on-hand inventory is known at all times.
2. Demand is random and stationary. That means that although we cannot predict the
value of demand, the expected value of demand over any time interval of fixed length
is constant. Assume that the expected demand rate is units per year.
3. There is a fixed positive lead time for placing an order.
4. The following costs are assumed:
Setup cost at $K per order.
Holding cost at $h per unit held per year.
Proportional order cost of $c per item.
Stock-out cost of $p per unit of unsatisfied demand. This is also called the shortage
cost or the penalty cost.
nah77856_ch05.indd 267
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 267
5.4
268 Chapter Five Inventory Control Subject to Uncertain Demand
Lot Size–Reorder Point Systems 267
FIGURE 5–5
Inventory
Changes in inventory
over time for
continuous-review
(Q, R) system
R
Q
Order
placed
Order Time
arrives
Describing Demand
In the newsvendor problem, the appropriate random variable is the demand during the
period. One period is the amount of time required to effect a change in the on-hand
inventory level. This is known as the response time of the system. In the context of our
current problem, the response time is the reorder lead time . Hence, the random
variable of interest is the demand during the lead time. We will assume that the demand
during the lead time is a continuous random variable D with probability density function (or pdf) f(x) and cumulative distribution function (or cdf) F(x). Let E(D) and
be the mean and the standard deviation of demand during the lead time.
var(D)
Decision Variables
There are two decision variables for this problem, Q and R, where Q is the lot size or order
quantity and R is the reorder level in units of inventory.2 Unlike the EOQ model, this
problem treats Q and R as independent decision variables. The policy is implemented as
follows: when the level of on-hand inventory reaches R, an order for Q units is placed that
will arrive in units of time. The operation of this system is pictured in Figure 5–5.
Derivation of the Expected Cost Function
The analytical approach we will use to solve this problem is, in principle, the same as
that used in the derivation of the newsvendor model. Namely, we will derive an expression
for the expected average annual cost in terms of the decision variables (Q, R) and
search for the optimal values of (Q, R) to minimize this cost.
2
When lead times are very long, it may happen that an order should be placed again before a prior order
arrives. In that case, the reorder decision variable R should be interpreted as the inventory position (onhand plus on-order) when a reorder is placed, rather than the inventory level.
nah77856_ch05.indd 268
17/12/14 7:10 PM
nah77856_ch05.qxd
268
1/9/08
11:42 AM
Page 268
Chapter Five Inventory Control Subject to Uncertain Demand
5.4 Lot Size–Reorder Point Systems 269
FIGURE 5–6
Inventory
Expected inventory
level for (Q, R)
inventory model
Slope = – Q
s
T = Q/
The Holding Cost
We assume that the mean rate of demand is units per year. The expected inventory
level varies linearly between s and Q s. We call s the safety stock; it is defined as the
expected level of on-hand inventory just before an order arrives, and is given by the
formula s R . The expected inventory level curve appears in Figure 5–6.
We estimate the holding cost from the average of the expected inventory curve. The
average of the function pictured in Figure 5–6 is s Q2 R Q2. An
important point to note here is that this computation is only an approximation. When
computing the average inventory level, we include both the cases when inventory is
positive and the cases when it is negative. However, the holding cost should not be
charged against the inventory level when it is negative, so that we are underestimating
the true value of expected holding cost. An exact expression for the true average
inventory is quite complex and has been derived only for certain specific demand distributions. In most real systems, however, the proportion of time spent out of stock is
generally small, so this approximation should be reasonably accurate.
Setup Cost
A cycle is defined as the time between the arrival of successive orders of size Q. Consistent with the notation used in Chapter 4, let T represent the expected cycle length.
Because the setup cost is incurred exactly once each cycle, we need to obtain an expression for the average length of a cycle in order to accurately estimate the setup cost
per unit of time.
There are a number of ways to derive an expression for the expected cycle length.
From Figure 5–6, we see that the distance between successive arrivals of orders is Q.
Another argument is the following. The expected demand during T is clearly T.
However, because the number of units that are entering inventory each cycle is Q and
there is conservation of units, the number of units demanded each cycle on average
also must be Q. Setting Q T and solving for T gives the same result.
nah77856_ch05.indd 269
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 269
270 Chapter Five Inventory Control Subject to Uncertain Demand
5.4
Lot Size–Reorder Point Systems 269
It follows, therefore, that the average setup cost incurred per unit time is KT KQ.
Penalty Cost
From Figure 5–5 we see that the only portion of the cycle during which the system is
exposed to shortages is between the time that an order is placed and the time that it arrives (the lead time). The number of units of excess demand is simply the amount by
which the demand over the lead time, D, exceeds the reorder level, R. It follows that the
expected number of shortages that occur in one cycle is given by the expression
E (max(D R, 0)) (x R) f(x) dx,
R
which is defined as n(R).
Note that this is essentially the same expression that we derived for the expected
number of stock-outs in the newsvendor model. As n(R) represents the expected number
of stock-outs incurred in a cycle, it follows that the expected number of stock-outs
incurred per unit of time is n(R)T n(R)Q.
Proportional Ordering Cost Component
Over a long period of time, the number of units that enter inventory and the number
that leave inventory must be the same. This means that every feasible policy will
necessarily order a number of units equal to the demand over any long interval of time.
That is, every feasible policy, on average, will replenish inventory at the rate of
demand. It follows that the expected proportional order cost per unit of time is c.
Because this term is independent of the decision variables Q and R, it does not affect
the optimization. We will henceforth ignore it.
It should be pointed out, however, that the proportional order cost will generally be
part of the optimization in an indirect way. The holding cost h is usually computed
by multiplying an appropriate value of the annual interest rate I by the value of the
item c. For convenience we use the symbol h to represent the holding cost, but keep in
mind that it could also be written in the form Ic.
The Cost Function
Define G(Q, R) as the expected average annual cost of holding, setup, and shortages.
Combining the expressions derived for each of these terms gives
G(Q, R) h(Q2 R ) KQ pn(R)Q.
The objective is to choose Q and R to minimize G(Q, R). We present the details of
the optimization in Appendix 5–C. As shown, the optimal solution is to iteratively
solve the two equations
Q
2[K pn(R)]
h
A
1 F(R) Qhp.
(1)
(2)
The solution procedure requires iterating between Equations (1) and (2) until two
successive values of Q and R are (essentially) the same. The procedure is started by
using Q0 EOQ (as defined in Chapter 4). One then finds R0 from Equation (2). That
value of R is used to compute n(R), which is substituted into Equation (1) to find Q1,
nah77856_ch05.indd 270
17/12/14 7:10 PM
nah77856_ch05.qxd
270
1/10/08
11:16 AM
Page 270
5.4 Lot Size–Reorder Point Systems 271
Chapter Five Inventory Control Subject to Uncertain Demand
which is then substituted into Equation (2) to find R1, and so on. Convergence generally occurs within two or three iterations. When units are integral, the computations
should be continued until successive values of both Q and R are within a single unit of
their previous values. When units are continuous, a convergence requirement of less
than one unit may be required depending upon the level of accuracy desired.
When the demand is normally distributed, n(R) is computed by using the standardized loss function. The standardized loss function L(z) is defined as
L(z) �
q
z
(t � z)(t) dt
where (t) is the standard normal density. If lead time demand is normal with mean and standard deviation , then it can be shown that
n(R) � L a
R�
b � L(z).
The standardized variate z is equal to (R � ). Calculations of the optimal policy
are carried out using Table A–4 at the back of this book.3
Example 5.4
Harvey’s Specialty Shop is a popular spot that specializes in international gourmet foods. One
of the items that Harvey sells is a popular mustard that he purchases from an English company.
The mustard costs Harvey $10 a jar and requires a six-month lead time for replenishment of
stock. Harvey uses a 20 percent annual interest rate to compute holding costs and estimates
that if a customer requests the mustard when he is out of stock, the loss-of-goodwill cost is $25
a jar. Bookkeeping expenses for placing an order amount to about $50. During the six-month
replenishment lead time, Harvey estimates that he sells an average of 100 jars, but there is
substantial variation from one six-month period to the next. He estimates that the standard
deviation of demand during each six-month period is 25. Assume that demand is described by
a normal distribution. How should Harvey control the replenishment of the mustard?
Solution
We wish to find the optimal values of the reorder point R and the lot size Q. In order to get the calculation started we need to find the EOQ. However, this requires knowledge of the annual rate of
demand, which does not seem to be specified. But notice that if the order lead time is six months
and the mean lead time demand is 100, that implies that the mean yearly demand is 200, giving
a value of � 200. It follows that the EOQ � 2Kh
� 1(2)(50)(200)(0.2)(10) � 100.
The next step is to find R0 from Equation (2). Substituting Q � 100, we obtain
1 � F(R0) � Q0hp � (100)(2)(25)(200) � .04.
From Table A–4 we find that the z value corresponding to a right tail of .04 is z � 1.75. Solving
R � z � gives R � (25)(1.75) � 100 � 144. Furthermore, z � 1.75 results in L(z) � 0.0162.
Hence, n(R) � L(z) � (25)(0.0162) � 0.405.
We can now find Q1 from Equation (1):
Q1 �
B
(2)(200)
[50 � (25)(0.405)] � 110.
2
This value of Q is compared with the previous one, which is 100. They are not close enough to
stop. Substituting Q � 110 into Equation (2) results in 1 � F(R1) � (110)(2)(25)(200) � .044.
Table A–4 now gives z � 1.70 and L(z) � 0.0183. Furthermore, R1 � (25)(1.70) � 100 � 143. We
3
Note that in rare cases Equations (1) and (2) may not be solvable. This can occur when the penalty cost
p is small compared to the holding cost h. The result is either diverging values of Q and R or the right
side of Equation (2) having a value exceeding 1 at some point in the calculation. In this case, the
recommended solution is to set Q � EOQ and R � R0 as long as the right side of Equation (2) is less
than 1. This often leads to negative safety stock, discussed later in this chapter.
nah77856_ch05.indd 271
17/12/14 7:10 PM
nah77856_ch05.qxd
1/10/08
11:16 AM
Page 271
272 Chapter Five Inventory Control Subject to Uncertain Demand
5.4
Lot Size–Reorder Point Systems 271
now obtain n(R1) � (25)(0.0183) � 0.4575, and Q2 � (200)[
50
25)(0.
� (4575)
] � 110.85 111. Substituting Q2 � 111 into Equation (2) gives 1 � F(R2) � .0444, z � 1.70, and R2 � R1 �
143. Because both Q2 and R2 are within one unit of Q1 and R1, we may terminate computations.
We conclude that the optimal values of Q and R are (Q, R) � (111, 143). Hence, each time
that Harvey’s inventory of this type of mustard hits 143 jars, he should place an order for 111 jars.
Example 5.4
(continued)
For the same example, determine the following:
Solution
1. The safety stock is s � R � � 143 � 100 � 43 jars.
2. We will compute average annual holding, setup, and penalty costs separately.
1. Safety stock.
2. The average annual holding, setup, and penalty costs associated with the inventory control
of the mustard.
3. The average time between placement of orders.
4. The proportion of order cycles in which no stock-outs occur.
5. The proportion of demands that are not met.
The holding cost is h[Q2 � R � ] � 2[1112 � 143 � 100] � $197 per year.
The setup cost is KQ � (50)(200)111 � $90.09 per year.
The stock-out cost is pn(R)Q � (25)(200)(0.4575)111 � $20.61 per year.
Hence, the total average annual cost associated with the inventory control of the mustard,
assuming an optimal control policy, is $307.70 per year.
3. T � Q � 111200 � 0.556 year � 6.7 months.
4. Here we need to compute the probability that no stock-out occurs in the lead time. This is
the same as the probability that the lead time demand does not exceed the reorder point.
We have PD � R � F(R) � 1 � .044 � .956. We conclude that there will be no stock-outs
in 95.6 percent of the order cycles.
5. The expected demand per cycle must be Q (see the argument in the derivation of the
expected setup cost). The expected number of stock-outs per cycle is n(R). Hence, the proportion of demands that stock out is n(R)Q � 0.4575111 � .004. Another way of stating
this result is that on average 99.6 percent of the demands are satisfied as they occur.
It should be noted in ending this section that Equations (1) and (2) are derived
under the assumption that all excess demand is back-ordered. That is, when an item is
demanded that is not immediately available, the demand is filled at a later time. However, in many competitive situations, such as retailing, a more accurate assumption is
that excess demand is lost. This case is known as lost sales. As long as the likelihood
of being out of stock is relatively small, Equations (1) and (2) will give adequate
solutions to both lost sales and back-order situations. If this is not the case, a slight
modification to Equation (2) is required. The lost-sales version of Equation (2) is
1 � F(R) � Qh(Qh � p).
(2�)
The effect of solving Equations (1) and (2�) simultaneously rather than Equations (1)
and (2) will be to increase the value of R slightly and decrease the value of Q slightly.
Inventory Level versus Inventory Position
An implicit assumption required in the analysis of (Q, R) policies was that each time an
order of Q units arrives, it increases the inventory level to a value greater than the reorder
point R. If this were not the case, one would never reach R, and one would never order
again. This will happen if the demand during the lead time exceeds Q, which is certainly
nah77856_ch05.indd 272
17/12/14 7:10 PM
Snapshot Application
nah77856_ch05.qxd
1/10/08
11:16 AM
Page 272
Inventory Management Software for
the Small Business
Inventory
The software for PC-based inventory management continues to grow at an increasing rate. As personal computers
have become more powerful, the inventory management
needs
of larger
businesses
beSubject
handled
by software
272 Chapter
Five
Inventory can
Control
to Uncertain
Demand
designed to run on PCs. One example is an inventory system
called5–7
inFlow, which can handle inventories stored at
FIGURE
multiple
locations.
It is relatively expensive at $299. A less
Inventory level versus
expensive
alternative
is Inventoria, which costs $79.99,
inventory position
which might be more appropriate for single location busiR�Q
nesses, such as one-off r­etail stores. Another choice for
small businesses is Small Business Inventory Control, which
has the advantage of allowing the user to enter items into
the system via barcode scanning. At $99, this package is
R choices include
also relatively inexpensive. Some other
iMagic Inventory, Inventory Power, Inventory Tracker Plus,
Inventory Executive System, and many others. Each of the
packages have certain advantages, including compatibility with other software systems, additional features such
as financial functions, single versus multiple l­ocations,
industry specific applications, etc. Of course, inventory
­
Order
management modules are also available in large
ERP
placed
­systems, such as those sold by United States-based Oracle
Corporation and German-based SAS.
An application that makes use of the most modern
hardware and software is ShopKeep POS. ShopKeep is
­designed to run on the Apple iPad, and all of the records
are stored in the cloud. This means that inventory data
can be accessed from any iPad in the system at any time.
One customer, The Bean, a small chain of coffee shops
based in New York City, ported their inventory control
function from a Windows-based system to ShopKeep,
running on iPads and iPhones. The system gives them the
opportunity to run reports from any iPad that’s connected
to the system. According to owner Ike Estavo, one of the
most useful functions of ShopKeep is identifying the most
profitable items per square foot of display space.
The personal computer has become ubiquitous since
the first edition of this book was published in 1989. The
Inventory position
huge growth in personal computers has been accompanied by an explosion in software choices for almost
every conceivable application. Inventory control is no
exception. A challenge for operations managers is to
choose wisely among the multitude of applications,
only a few of which are mentioned here. One must be
Inventory level
able to pick the level of generality and complexity that’s
appropriate to the particular application. But no matter
what software package is chosen, it is vitally ­important
to understand the basic principles of inventory manTime
Order
agement discussed here. Hoping that
the software will
placed
solve one’s problem without really understanding the
key elements of that problem is a big mistake.
Source: The information discussed here was obtained from
the websites of the vendors mentioned.
possible. On the surface, this seems like a serious flaw in the approach, but it isn’t. To
avoid this problem, one bases the reorder decision on the inventory position rather than
on the inventory level. The inventory position is defined as the total stock on hand plus
on order. The inventory position varies from R to R � Q as shown in Figure 5–7.
Figure 5–7
5.5
InventorySERVICE
level versus LEVELS IN (Q, R) SYSTEMS
AlthoughR theQinventory model described in Section 5.4 is quite realistic for describing
position
many real systems, managers often have a difficultInventory
time determining
an exact value for
the stock-out cost p. In many cases, the stock-out cost includes intangible components
such as loss of goodwill and potential delays to other parts of the system. A common
substitute forRa stock-out cost is a service level. Although there are a number of different
definitions of service, it generally refers to the probability that a demand or a collection
of demands is met. Service levels can be appliedInventory
in both level
periodic review and (Q, R)
systems. The application of service levels to periodic-review systems will be discussed
in Section 5.6. Service levels for continuous-review systems are considered here.
Two types of service are considered, labeled Type 1 and Type 2, respectively.
Inventory
inventory position
Type 1 Service
Order
placed
Order
placed
Time
In this case we specify the probability of not stocking out in the lead time. We will use
the symbol to represent this probability. As specification of completely determines
the value of R, the computation of R and Q can be decoupled. The computation of the
optimal (Q, R) values subject to a Type 1 service constraint is very straightforward.
a. Determine R to satisfy the equation F(R) � .
b. Set Q � EOQ.
nah77856_ch05.indd 273
Interpret as the proportion of cycles in which no stock-out occurs. A Type 1 service objective is appropriate when a shortage occurrence has the same consequence
17/12/14
independent of its time or amount. One example would be where a production line
7:10 PM
possible. On the surface, this seems like a serious flaw in the approach, but it isn’t. To
avoid this problem, one bases the reorder decision on the inventory position rather than
on the inventory level. The inventory position is defined as the total stock on hand plus
on Control
order.Subject
The inventory
274 Chapter Five Inventory
to Uncertainposition
Demand varies from R to R � Q as shown in Figure 5–7.
5.5
SERVICE LEVELS IN (Q, R) SYSTEMS
Although the inventory model described in Section 5.4 is quite realistic for describing
many real systems, managers often have a difficult time determining an exact value for
the stock-out cost p. In many cases, the stock-out cost includes intangible components
such as loss of goodwill and potential delays to other parts of the system. A common
substitute for a stock-out cost is a service level. Although there are a number of different
definitions of service, it generally refers to the probability that a demand or a collection
of demands is met. Service levels can be applied in both periodic review and (Q, R)
systems. The application of service levels to periodic-review systems will be discussed
in Section 5.6. Service levels for continuous-review systems are considered here.
Two types of service are considered, labeled Type 1 and Type 2, respectively.
Type 1 Service
nah77856_ch05.qxd
1/10/08
11:16
In this case we specify the probability of not stocking out in the lead time. We will use
the symbol to represent this probability. As specification of completely determines
the value of R, the computation of R and Q can be decoupled. The computation of the
optimal
AM
Page(Q,
273R) values subject to a Type 1 service constraint is very straightforward.
a. Determine R to satisfy the equation F(R) � .
b. Set Q � EOQ.
Interpret as the proportion of cycles in which no stock-out occurs. A Type 1 service objective is appropriate when a shortage occurrence has the same consequence
5.5 Service
Levels ina(Q,
R) Systems line
273
independent of its time or amount. One example would
be where
production
is stopped whether 1 unit or 100 units are short. However, Type 1 service is not how
service is interpreted in most applications. Usually when we say we would like to
provide 95 percent service, we mean that we would like to be able to fill 95 percent
of the demands when they occur, not fill all the demands in 95 percent of the order
cycles. Also, as different items have different cycle lengths, this measure will not be
consistent among different products, making the proper choice of difficult.
Type 2 Service
Type 2 service measures the proportion of demands that are met from stock. We will
use the symbol to represent this proportion. As we saw in part 5 of Example 5.4,
n(R)Q is the average fraction of demands that stock out each cycle. Hence, specification of results in the constraint n(R)Q � 1 � .
This constraint is more complex than the one arising from Type 1 service, as it involves both Q and R. It turns out that although the EOQ is not optimal in this case, it
usually gives pretty good results. If we use the EOQ to estimate the lot size, then we
would find R to solve n(R) � EOQ(1 � ).
Example 5.5
Consider again Harvey’s Specialty Shop, described in Example 5.4. Harvey feels uncomfortable
with the assumption that the stock-out cost is $25 and decides to use a service level criterion
instead. Suppose that he chooses to use a 98 percent service objective.
1. Type 1 service. If we assume an of .98, then we find R to solve F(R) � 0.98. From Table A–1
or A–4, we obtain z � 2.05. Setting R � z � gives R � 151.
2. Type 2 service. Here � 0.98. We are required to solve the equation
n(R) � EOQ(1 � ),
which is equivalent to
L(z) � EOQ(1 � ).
Substituting EOQ � 100 and � .98, we obtain
L(z) � (100)(0.02)25 � 0.08.
From Table A–4 of the unit normal partial expectations, we obtain z � 1.02. Setting R � z � gives R � 126. Notice that the same values of and give considerably different values of R.
nah77856_ch05.indd 274
In order to understand more clearly the difference between these two measures of
17/12/14 7:10 PM
Example
Example 5.5
5.5
Example 5.5
usually gives pretty good results. If we use the EOQ to estimate the lot size, then we
would
R
solve
n(R)
�
�
).
usually
gives
pretty
If we
would find
find
R to
to
solvegood
n(R)results.
� EOQ(1
EOQ(1
� use
). the EOQ to estimate the lot size, then we
would find R to solve n(R) � EOQ(1 � ).
Consider again Harvey’s Specialty Shop, described in Example 5.4. Harvey feels uncomfortable
Consider again Harvey’s Specialty Shop, described in Example 5.4. Harvey feels uncomfortable
with the assumption
thatSpecialty
the stock-out
is $25inand
decides
toHarvey
use a service
level criterion
Consider
again Harvey’s
Shop, cost
described
Example
5.4.
feels uncomfortable
with the assumption
that the stock-out
cost
is $25 and
decides
to use a service
level criterion
instead.
Suppose
thatthat
he chooses
to usecost
a 98ispercent
service
objective.
with
the
assumption
the
stock-out
$25
and
decides
to
use
a
service
level criterion
instead. Suppose that he chooses to use a 98 percent service objective.
instead. Suppose that he chooses to use a 98 percent service objective.
1. Type 1 service. If we assume an of .98, then we find R to solve F(R) � 0.98. From Table A–1
1. Type 1 service. If we assume an of .98, then we find R to solve F(R) � 0.98. From Table A–1
or A–4,
we obtain
� 2.05.an
Setting
R �then
z � gives
�solve
151.
1. Type
1 service.
If wezz assume
of .98,
find RR
F(R)
� 0.98.
Table A–1
5.5
Service
Levels
in (Q,From
R) Systems 275
or A–4,
we obtain
� 2.05. Setting
R � z �we
gives
Rto
� 151.
A–4,
we obtain
��
2.05.
Setting
R�
z � togives
� 151.
2. or
Type
2 service.
Herez 0.98.
We are
required
solveR the
equation
2. Type 2 service. Here � 0.98. We are required to solve the equation
2. Type 2 service. Here � 0.98. We are required to solve the equation
n(R) � EOQ(1 � ),
n(R) � EOQ(1 � ),
n(R) � EOQ(1 � ),
which is equivalent to
which is equivalent to
L(z) � EOQ(1 � ).
which is equivalent to
L(z) � EOQ(1 � ).
L(z) � EOQ(1 � ).
Substituting EOQ � 100 and � .98, we obtain
Substituting EOQ � 100 and � .98, we obtain
Substituting EOQ � 100 andL(z)
�
we obtain � 0.08.
� .98,
(100)(0.02)25
L(z) � (100)(0.02)25 � 0.08.
L(z) � (100)(0.02)25 � 0.08.
From Table A–4 of the unit normal partial expectations, we obtain z � 1.02. Setting R � z � From Table A–4 of the unit normal partial expectations, we obtain z � 1.02. Setting R � z � gives RTable
� 126.
samepartial
values of and give
considerably
different
values
of R.
From
A–4Notice
of thethat
unit the
normal
we obtain
z � 1.02.
Setting
R � z
�
gives R � 126.
Notice
that
the
same valuesexpectations,
of and give
considerably
different
values
of R.
gives R � 126. Notice that the same values of and give considerably different values of R.
In
In order
order to
to understand
understand more
more clearly
clearly the
the difference
difference between
between these
these two
two measures
measures of
of
service,
consider
the
following
example.
Suppose
that
we
have
tracked
the
demands
In
order
to
understand
more
clearly
the
difference
between
these
two
measures
of
service, consider the following example. Suppose that we have tracked the
demands
and
stock-outs
over
10
consecutive
order
cycles
with
the
following
results:
service,
consider
the
following
example.
Suppose
that
we
have
tracked
the
demands
and stock-outs over 10 consecutive order cycles with the following results:
and stock-outs over 10 consecutive order cycles with the following results:
nah77856_ch05.qxd
274
1/10/08
11:16 AM
Order Cycle
Order Cycle
Order Cycle
1
1
2
1
2
3
2
3
4
3
4
5
4
5
6
5
6
Page 274 6
7
7
8
7
8
9
8
9
10
9
10
10
Demand
Demand
Demand
180
180
75
180
75
235
75
235
140
235
140
180
140
180
200
180
200
150
200
150
90
150
90
160
90
160
40
160
40
40
Stock-Outs
Stock-Outs
Stock-Outs
0
0
0
0
45
0
45
0
45
0
0
0
10
0
10
0
10
0
0
0
0
0
0
0
0
0
Based
Based on
on aa Type
Type 11 measure
measure of
of service,
service, we
we find
find that
that the
the fraction
fraction of
of periods
periods in
in which
which
there is no stock-out is 810 � 80 percent. That is, the probability that all the demands are
met
order
cycle
is
0.8,
based
on
However,
the
therein
stock-out
is 810
percent.
is, theobservations.
probability that
all the demands
met
inis aanosingle
single
order
cycle�
is80
0.8,
basedThat
on these
these
observations.
However,
the Type
Typeare22
service
herecycle
is considerably
better.
In thisobservations.
example, theHowever,
total number
of de-2
met
in aprovided
single order
is 0.8, based
on these
the Type
mands over the 10 periods is 1,450 (the sum of the numbers in the second column), and
the total number of demands that result in a stock-out is 55. Hence, the number of satisfied demands is 1,450 � 55 � 1,395. The proportion of satisfied demands is 1,3951,450
� .9621, or roughly 96 percent.
The term fill rate is often used to describe Type 2 service, and is generally what most
managers mean by service. (The fill rate in this example is 96 percent.) We saw in the
example that there is a significant difference between the proportion of cycles in which
all demands are satisfied (Type 1 service) and the fill rate (Type 2 service). Even though
it is easier to determine the best operating policy satisfying a Type 1 service objective,
this policy will not accurately approximate a Type 2 service objective and should not
be used in place of it.
Chapter Five Inventory
Control
to Uncertain
Demand
there
is noSubject
stock-out
is
810
� 80
is, the
probability
thatof
allperiods
the demands
are
Based
on
a Type
1 measure
ofpercent.
service,That
we find
that
the fraction
in which
Optimal (Q, R) Policies Subject to Type 2 Constraint
Using the EOQ value to estimate the lot size gives reasonably accurate results when
using a fill rate constraint, but the EOQ value is only an approximation of the optimal
lot size. A more accurate value of Q can be obtained as follows. Consider the pair of
equations (1) and (2) we solved for the optimal values of Q and R when a stock-out cost
was present. Solving for p in Equation (2) gives
p � Qh[(1 � F(R))],
which now can be substituted for p in Equation (1), resulting in
Q�
2{K � Qhn(R)>[(1 � F(R))]}
A
h
,
which is a quadratic equation in Q. It can be shown that the positive root of this
equation is
nah77856_ch05.indd 275
17/12/14 7:10 PM
nah77856_ch05.qxd
274
1/10/08
Optimal
(Q,
11:16 AM
Page 274
R) Policies Subject to Type 2 Constraint
Using the EOQ value to estimate the lot size gives reasonably accurate results when
using a fill rate constraint, but the EOQ value is only an approximation of the optimal
lot size. A more accurate value of Q can be obtained as follows. Consider the pair of
equations (1) and (2) we solved for the optimal values of Q and R when a stock-out cost
was present. Solving for p in Equation (2) gives
Chapter
Five Inventory
Control Control
Subject to
Uncertain
DemandDemand
276 Chapter
Five Inventory
Subject
to Uncertain
p � Qh[(1 � F(R))],
which
now can
forbetter.
p in Equation
(1), resulting
in number of deservice
provided
herebeissubstituted
considerably
In this example,
the total
mands over the 10 periods is 1,450 (the sum of the numbers in the second column), and
the total number of demands that result
in a�stock-out
is 55.
the number of satis2{K
Qhn(R)>[(1
�Hence,
F(R))]}
� The proportion of satisfied demands
, is 1,3951,450
fied demands is 1,450 � 55 �Q1,395.
h
� .9621, or roughly 96 percent. A
The term fill rate is often used to describe Type 2 service, and is generally what most
which
is abyquadratic
equation
in Q.
It can
be shown
the positive
managers
mean
service. (The
fill rate
in this
example
is 96 that
percent.)
We saw root
in theof this
equation
is
example that there is a significant difference between the proportion of cycles in which
all demands are satisfied (Type 1 service)
fill rate (Type
2 service).
Even though
2
n(R) and the2K
n(R)
� operating �
it is easier to determine the Q
best
policy satisfying
� a a Type 1bservice objective, (3)
1 � F(R) A
1 �objective
F(R) and should not
this policy will not accurately approximate a Typeh 2 service
be used in
place of(3)
it. will be called the SOQ formula (for service level order quantity).4 This
Equation
equation
solved
simultaneously
Optimal
(Q,isR)
Policies
Subjectwith
to Type 2 Constraint
nah77856_ch05.qxd
1/9/08
(1 �reasonably
)Q
Using the EOQ value to estimate the lotn(R)
size�gives
accurate results when (4)
using a fill rate constraint, but the EOQ value is only an approximation of the optimal
to obtain optimal values of (Q, R) satisfying a Type 2 service constraint.
lot size. A more accurate value of Q can be obtained as follows. Consider the pair of
The reader should note that the version of Equation (4) used in the calculations is in
equations (1) and (2) we solved for the optimal values of Q and R when a stock-out cost
terms
of 275
the standardized variate z and is given by
11:42 AM Page
was present. Solving
for p in Equation (2) gives
L(z) � (1 � )Q.
p � Qh[(1 � F(R))],
The solution procedure is essentially the same as that required to solve Equations (1)
which now can be substituted for p in Equation (1), resulting in
and (2) simultaneously. Start with Q0 � EOQ, find R0 from (4), use R0 in (3) to find Q1,
and so on, and stop when two successive values of Q and R are sufficiently close (within
2{Kproblems).
� Qhn(R)>[(1 � F(R))]}
5.5 Service Levels in (Q, R) Systems 275
one unit is sufficient for most
Q�
,
A
h
Returning to Example 5.5, Q0 � 100 and R0 � 126. Furthermore, n(R0) � (0.02)(100) � 2. Using
Example 5.5
� 1.02
gives
1�
F(Rcould
0.154.
the
calculations,
formula
also
bywith
more
conventional
multiplier
techniques.
0) � have
is
aSOQ
quadratic
equation
in been
Q.Continuing
Itderived
can be
shown
that theLagrange
positive
root of
this
(continued) whichz4 The
We include
this derivation to demonstrate the relationship between the fill rate objective and the
equation
is
2
stock-out cost model.
Q�
Q1 �
2
2
� (100)2 �
0.1542KB
0.154
2
n(R)
n(R)
�
�a
b
� 114.
1 � F(R)
A h
1 � F(R)
(3)
Solving Equation (4) gives n(R1) � (114)(0.02), which is equivalent to
Equation (3) will be called the SOQ formula (for service level order quantity).4 This
� (114)(0.02)25 � 0.0912.
equation is solved simultaneously L(z)
with
From Table A–4, z � 0.95, son(R)
that� (1 � )Q
(4)
1 � F(R
1) � 0.171
to obtain optimal values of (Q, R) satisfying
a Type
2 service constraint.
Theand
reader should note that the version of Equation (4) used in the calculations is in
terms of the standardized variate z and is given by
R1 � z � � 124.
L(z) � (1 � )Q.
Carrying the computation one more step gives Q2 � 114 and R2 � 124. As both Q and R are
Thewithin
solution
essentially
the we
same
as that computations.
required to solve
Equations
(1) that
one procedure
unit of theirisprevious
values,
terminate
Hence,
we conclude
and (2)
simultaneously.
Start
with
Q0 � EOQ,
R0 from
(4),
use R0 in
Q1, 124).
the
optimal values of
Q and
R satisfying
a 98find
percent
fill rate
constraint
are(3)
(Q,toR)find
� (114,
Consider
thewhen
cost error
resulting from
the EOQ
the SOQ. close
In order
to compare
and so on,
and stop
two successive
values
of Qsubstituted
and R are for
sufficiently
(within
these
policies, wefor
compute
the average annual holding and setup costs (notice that there is no
one unit
is sufficient
most problems).
stock-out cost) for the policies (Q, R) � (100, 126) and (Q, R) � (114, 124).
Recall the formulas for average annual holding and setup costs.
4
The SOQ formula also could have been derived
by more
Holding
cost conventional
� h(Q2 � Lagrange
R � ). multiplier techniques.
We include this derivation to demonstrate the relationship between the fill rate objective and the
Setup cost � KQ.
stock-out cost model.
For (100, 126):
Total � $252
Total � $250
Holding cost � 2(1002 � 126 � 100) � $152
Setup cost � (50)(200)100
� $100
For (114, 124):
Holding cost � 2(1142 � 124 � 100) � $162
Setup cost � (50)(200)114
� $88
nah77856_ch05.indd 276
17/12/14 7:10 PM
and
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 276
R1 � z � � 124.
Carrying the computation one more step gives Q2 � 114 and R2 � 124. As both Q and R are
within one unit of their previous values, we terminate computations. Hence, we conclude that
the optimal values of Q and R satisfying a 98 percent fill rate constraint are (Q, R) � (114, 124).
Consider the cost error resulting from the EOQ substituted for the SOQ. In order to compare
276 Chapter Five Inventory
Control
Subjectwe
to Uncertain
Service
Levels
in (Q,that
R) Systems 277
these
policies,
computeDemand
the average annual holding and5.5
setup
costs
(notice
there is no
stock-out cost) for the policies (Q, R) � (100, 126) and (Q, R) � (114, 124).
Recall the
average
annual holding and setup costs.
Scaling
offormulas
Lead for
Time
Demand
� h(Q2
R � time
). was given. However, in
In all previous examples theHolding
demandcost
during
the�lead
most applications demand would
becost
forecast
on a periodic basis, such as monthly.
Setup
� KQ.
In
such
cases
one
would
need
to
convert
the
demand
distribution to correspond to the
For (100, 126):
lead time.
Holding
cost �follow
2(1002
� 126 �
100) � $152
Assume that
demands
a normal
distribution.
Because
sums
of independent
Total
� $252
Setup
cost
�
(50)(200)100
� $100the form of the distribution of
normal random variables are also normally distributed,
lead
time124):
demand is normal. Hence, all that remains is to determine the mean and the
For (114,
standard deviation. Let the periodic demand have mean and standard deviation ,
cost � 2(1142 � 124 � 100) � $162
and let beHolding
the lead
time in periods. As both the means and
variances (not
Totalthe
� $250
Setup cost
(50)(200)114
� $88
standard deviations)
are�additive,
the mean demand
during lead time is � and
Hence,
theinstandard
the
variance
demand
during lead
2.
We see
that theofEOQ
approximation
gives time
costs is
close
to the
optimal
this case.deviation of
demand during lead time is � (although the square root may not always be
5
appropriate).
Imputed Shortage Cost
Example 5.6
Solution
nah77856_ch05.qxd
1/9/08
Example 5.5
(continued)
276 Chapter Five
Example 5.6
Solution
Consider the solutions that we obtained for (Q, R) in Example 5.5 when we used a
Weekly
for a certain
sparkFor
plug
in a local
repair of
shop
is normally
service demand
level criterion
rathertype
thanofaautomotive
shortage cost.
a Type
2 service
�
0.98 we
distributed
with
mean (114,
34 and
standard
deviation
12. Procurement
lead time this
is sixsolution
weeks.
obtained the
solution
124).
Although
no shortage
cost was specified,
Determine the lead time demand distribution.
clearly corresponds to some value of p. That is, there is some value of p such that the
policy
(114, 124)
satisfies
Equations
and (2).
This particular
value(34)(6)
of p is�known
as
The demand
over the
lead time
is also(1)
normally
distributed
with mean
204 and
standard
deviation
(12)6
� 29.39. These would be the values of and that one would use
the imputed
shortage
cost.
for The
all remaining
imputed calculations.
shortage cost is easy to find. One solves for p in Equation (2) to obtain
p
�
Qh[(1
�
F(R))].
The imputed shortage cost is a useful way to determine whether
11:42 AM Page 276
the
value
chosen
for
the
service level
is appropriate.
Estimating Sigma When
Inventory
Control
and Forecasting Are Linked
Consider
Usingthat
a value
� .98 (Type
1 service),over
we obtained
Thus
far again
in thisHarvey’s
chapterSpecialty
we haveShop.
assumed
the of
distribution
of demand
the lead
the policy
(100, 151).
The imputed
shortage acost
is pfor
� (100)(2)[(0.02)(200)]
time
is known.
In practice,
one assumes
form
the distribution, but �its$50.
parameters
Using
a value of from
� 0.98
service) we aobtained
policy (114,
this demand
case the
must
be estimated
real(Type
data.2 Assuming
normal the
distribution
for124).
leadIntime
Inventory
Controlcost
Subject
toshortage
UncertainisDemand
imputed
of
p
�
(114)(2)[(0.171)(200)]
�
$6.67.
(which is the most common assumption), one needs to estimate the mean and the standard
deviation. When a complete history of past data is available, the standard statistical
Scaling
of the
Lead
Time
Demand
estimates for
mean
and the
standard deviation (i.e., those suggested in Section 5.1)
In
previous
examples
the demand
duringdothe
time
was
given.
However,
in
areall
fine.
However,
most forecasting
schemes
notlead
use all
past
data.
Moving
averages
most
applications
would
forecast on
a periodic
basis,
such asweights
monthly.
use only
the past Ndemand
data values
and be
exponential
smoothing
places
declining
on
In
such
cases one would need to convert the demand distribution to correspond to the
past
data.
lead
Intime.
these cases, it is unclear exactly what are the right estimators for the mean and the
Assume
that demands
follow
a normal
distribution.
sums
independent
standard
deviation
of demand.
This
issue was
discussed Because
in Section
2.13.ofThe
best estinormal
normally
distributed,
form
of the
distribution
of
mate of random
the meanvariables
is simplyare
thealso
forecast
of demand
for thethe
next
period.
For
the variance,
lead
time demand
normal. for
Hence,
all that remains
is toerror.
determine
the mean
one should
use theisestimator
the variance
of forecast
The rationale
forand
thisthe
is
standard
deviation.The
Let variance
the periodic
demanderror
haveand
mean
and standard
deviation
,
rarely understood.
of forecast
the variance
of demand
are not
and
let thing.
be theThis
lead
time
in periods.
As both
means2–A.
and the variances (not
the same
was
established
rigorously
in the
Appendix
standard
are to
additive,
mean deviation
demand during
lead error
time is
� and
Why isdeviations)
it appropriate
use the the
standard
of forecast
to estimate
?
2
. Hence,
the standard
of
the
is using
The variance
reason is of
thatdemand
it is theduring
forecastlead
thattime
we are
to estimate
demand. deviation
Safety stock
(althoughdemand.
the square
root maythe
notvariance
always be
demand
lead
time is
� in
is held toduring
protect
against
errors
forecasting
In general,
of
5
appropriate).
forecast error will be higher than the variance of demand. This is the result of the
Weekly demand for a certain type of automotive spark plug in a local repair shop is normally
5
Often in practice
turns out
is moredeviation
variation in
theProcurement
demand process
than
is described
by a
distributed
with itmean
34 that
and there
standard
12.
lead
time
is six weeks.
pure normal the
distribution.
Fordemand
that reason
the standard deviation of demand is generally expressed in the
Determine
lead time
distribution.
form q where the correct value of q, generally between 0.5 and 1, must be determined for each item
or group
of items
by an
of historical
The
demand
over
theanalysis
lead time
is alsodata.
normally distributed with mean (34)(6) � 204 and
standard deviation (12)6
� 29.39. These would be the values of and that one would use
for all remaining calculations.
Estimating Sigma When Inventory Control
and Forecasting Are Linked
nah77856_ch05.indd 277
Thus far in this chapter we have assumed that the distribution of demand over the lead
time is known. In practice, one assumes a form for the distribution, but its parameters
17/12/14 7:10 PM
lead time demand is normal. Hence, all that remains is to determine the mean and the
standard deviation. Let the periodic demand have mean and standard deviation ,
and let be the lead time in periods. As both the means and the variances (not
standard deviations) are additive, the mean demand during lead time is � and
the variance of demand during lead time is 2. Hence, the standard deviation of
demand during lead time is � (although the square root may not always be
5
appropriate).
278 Chapter Five Inventory
Control Subject to Uncertain Demand
Example 5.6
Weekly demand for a certain type of automotive spark plug in a local repair shop is normally
distributed with mean 34 and standard deviation 12. Procurement lead time is six weeks.
Determine the lead time demand distribution.
Solution
The demand over the lead time is also normally distributed with mean (34)(6) � 204 and
standard deviation (12)6
� 29.39. These would be the values of and that one would use
for all remaining calculations.
Estimating Sigma When Inventory Control
and Forecasting Are Linked
nah77856_ch05.qxd
1/9/08
11:42
Thus far in this chapter we have assumed that the distribution of demand over the lead
time is known. In practice, one assumes a form for the distribution, but its parameters
must be estimated from real data. Assuming a normal distribution for lead time demand
(which is the most common assumption), one needs to estimate the mean and the standard
deviation. When a complete history of past data is available, the standard statistical
estimates for the mean and the standard deviation (i.e., those suggested in Section 5.1)
are fine. However, most forecasting schemes do not use all past data. Moving averages
use only the past N data values and exponential smoothing places declining weights on
past data.
In these cases, it is unclear exactly what are the right estimators for the mean and the
standard deviation of demand. This issue was discussed in Section 2.13. The best estiof the
AMmate
Page
277mean is simply the forecast of demand for the next period. For the variance,
one should use the estimator for the variance of forecast error. The rationale for this is
rarely understood. The variance of forecast error and the variance of demand are not
the same thing. This was established rigorously in Appendix 2–A.
Why is it appropriate to use the standard deviation of forecast error to estimate ?
The reason is that it is the forecast that we are using to estimate demand. Safety stock
5.5 Service
Levels in (Q,
Systems 277
is held to protect against errors in forecasting demand.
In general,
theR)variance
of
forecast error will be higher than the variance of demand. This is the result of the
additional sampling error introduced by a forecasting scheme that uses only a portion
of past data.6
5
Often
in practice
out that
is more variation
in the
process than
is described
by a
R. G.
Brownit turns
(1959)
wasthere
apparently
the first
todemand
recommend
using
the standard
pure normalof
distribution.
For
thatin
reason
thestock
standard
deviation of demand
is generally
expressed in the
deviation
forecast
error
safety
calculations.
The
method
he
recommended,
form q where the correct value of q, generally between 0.5 and 1, must be determined for each item
which
is still in widespread use today, is to track the MAD (mean absolute deviation)
or group of items by an analysis of historical data.
of forecast error using the formula
MADt � � MADt�1 � (1 � �)Ft � Dt where Ft is the forecast of demand at time t and Dt is the actual observed demand at
time t. The estimator for the standard deviation of forecast error at time t is 1.25 MADt.
While this method is very popular in commercial inventory control systems, apparently few realize that by using this approach they are estimating not demand variance
but forecast error variance, and that these are not the same quantities.
*Lead Time Variability
6
Thus
far,where
we have
assumeddemand
that the
leadistime
� is a known
However,
lead time
In cases
the underlying
process
nonstationary,
that is,constant.
where there
is trend or seasonality,
uncertainty
common
practice.
Forcould
example,
required
to transport
the
variance of is
demand
due to in
systematic
changes
be higherthe
thantime
the variance
of forecast
error. All
of
our analysis assumes
stationary
(constant
mean and
demand
patterns,
however.
commodities,
such as
oil, that
are shipped
byvariance)
sea depends
upon
weather
conditions. In
nah77856_ch05.indd 278
general, it is very difficult to incorporate the variability of lead time into the calculation of optimal inventory policies. The problem is that if we assume that successive
lead times are independent random variables, then it is possible for lead times to cross;
that is, two successive orders would not necessarily be received in the same sequence
that they were placed.
Order crossing is unlikely when a single supplier is used. If we are willing to make
the simultaneous assumptions that orders do not cross and that successive lead times
are independent, the variability of lead time can be easily incorporated into the analysis. Suppose that the lead time � is a random variable with mean �� and variance � 2�.
Furthermore, suppose that demand in any time t has mean �t and variance � 2t. Then it
17/12/14 7:10 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 277
MADt � � MADt�1 � (1 � �)Ft � Dt where Ft is the forecast of demand at time t and Dt is the actual observed demand at
time t. The estimator for the standard deviation of forecast error at time t is 1.25 MADt.
While this method is very popular in commercial inventory control systems, apparently few realize that by using this approach they are estimating not demand variance
5.5 Service Levels in (Q, R) Systems 277
but forecast error variance, and that these are not the same
quantities.
5.5 Service
Levels in (Q, R) Systems 279
additional sampling error introduced by a forecasting scheme that uses only a portion
6
*Lead
Time
Variability
of
past data.
Thus
far,
we
have
assumed
thatapparently
the lead time
is a known
constant. using
However,
time
R. G. Brown (1959)
was
the� first
to recommend
the lead
standard
uncertainty
is
common
in
practice.
For
example,
the
time
required
to
transport
deviation of forecast error in safety stock calculations. The method he recommended,
commodities,
such
as oil, that
aretoday,
shipped
sea depends
weather
conditions.
In
which
is still in
widespread
use
is tobytrack
the MADupon
(mean
absolute
deviation)
general,
it
is
very
difficult
to
incorporate
the
variability
of
lead
time
into
the
calculaof forecast error using the formula
tion of optimal inventory policies. The problem is that if we assume that successive
lead times are independent
variables,
is possible
(1 �it �)F
MADrandom
t � � MAD
t�1 �then
t � Dt for lead times to cross;
that is, two successive orders would not necessarily be received in the same sequence
that they
is theplaced.
forecast of demand at time t and Dt is the actual observed demand at
where
Ft were
Order
is unlikely
when a deviation
single supplier
is used.
If we
are willing
maket.
time
t. Thecrossing
estimator
for the standard
of forecast
error
at time
t is 1.25toMAD
the simultaneous
that orders
do not cross
and that
successive
leadappartimes
While
this methodassumptions
is very popular
in commercial
inventory
control
systems,
are independent,
of lead
time can
easily
incorporated
into the
analyently
few realize the
thatvariability
by using this
approach
theybeare
estimating
not demand
variance
sis. forecast
Supposeerror
that the
lead time
is athese
random
variable
withquantities.
mean �� and variance � 2�.
but
variance,
and �that
are not
the same
Furthermore, suppose that demand in any time t has mean �t and variance � 2t. Then it
can be shown
the demand during lead time has mean and variance7
*Lead
Timethat
Variability
� ��
Thus far, we have assumed that the�lead
time
�, � is a known constant. However, lead time
2
uncertainty is common in practice.
For
� �2� 2�. the time required to transport
� � ���2example,
commodities, such as oil, that are shipped by sea depends upon weather conditions. In
general,
it isthe
very
difficult
to incorporate
the variability
lead time
into the
Example 5.7
Harvey Gold,
owner
of Harvey’s
Specialty Shop,
orders an of
unusual
olive from
the calculaisland of
tion
of optimal
inventory
is that
if we
assume that
successive
Santorini,
off the Greek
coast.policies.
Over theThe
years,problem
Harvey has
noticed
considerable
variability
in the
lead
it is the
possible
cross;
time times
it takesare
to independent
receive ordersrandom
of thesevariables,
olives. On then
average,
order for
leadlead
timetimes
is fourtomonths
and the
standard
deviation
is six weeks
Monthly
demand for
normally
that
is, two
successive
orders
would(1.5
notmonths).
necessarily
be received
in the olives
same issequence
nah77856_ch05_248-310.qxd 2/17/10
PMplaced.
Page
distributed
with
mean
15 278
(jars) and standard deviation 6.
that
they5:15
were
Settingcrossing
�� � 4, �is
1.5, � �when
15, and
� � 6,supplier
we obtainis used. If we are willing to make
��
Order
unlikely
a single
the simultaneous assumptions that
not �
cross
(4)(15)
60, and that successive lead times
� � orders
�� � � do
are independent, the2 variability
of
time can be easily incorporated into the analy2
2 lead
2
� � � � � � � � (4)(36) � (225)(2.25) � 650.25.
sis. Suppose that the lead� time � is �a random variable with mean �� and variance � 2�.
One would suppose
proceed with
the calculations
optimal
policies
using ��2�
60 and
t. Then
it
Furthermore,
that demand
in any of
time
t has inventory
mean �t and
variance
2 Control Subject to Uncertain Demand
278 Chapter Five Inventory
7
�
�
650.25
as
the
mean
and
the
variance
of
lead
time
demand.
can be shown that the demand during lead time has mean and variance
� � ��
�,
Calculations
in Exceldemand process
In cases where the underlying
is nonstationary,
that is, where there is trend or
6
2
Example 5.7
2
2 2
seasonality,
the variance
of function,
demand dueL(z),
to systematic
changes
higherTo
than
of
�
��be
� computed
� � � could
�
�. inbe
The
standardized
loss
can
Excel.
dothe
so,variance
write L(z)
in
forecast error. All of our analysis assumes stationary (constant mean and variance) demand patterns,
the
following
way:
however.
Harvey
Gold, the
Shop, orders an unusual olive from the island of
� owner of Harvey’s Specialty
�
7
Hadley and
Whitin (1963),coast.
p. 153.
Santorini,
years,dtHarvey
noticed�considerable
L(z) off
� the(tGreek
� z)(t) dtOver
� thet(t)
� z(1has
� �(z))
(z) � z(1 variability
� �(z)).in the
z olives. On average, the order lead time is four months
time it takes toz receive orders of these
and
standard
deviation
six weeks
(1.5 months).
demand for
the olives isfunction,
normally
Thethe
first
equality
resultsisfrom
the definition
of Monthly
the cumulative
distribution
distributed
with
mean
15
(jars)
and
standard
deviation
6.
and the second equality is a consequence of a well-known property of the standard norSetting �� � 4, �� � 1.5, � � 15, and � � 6, we obtain
mal distribution (see, for example, Hadley and Whitin, 1963, p. 444). To program this
formula into Excel, use the definition
standard
normal density function, which is
� (4)(15)
� 60,
� � ��of� the
� (225)(2.25) � 650.25.
� 2 � �� � 2 � �2 � 2� � (4)(36)
1
(z) �
exp (�0.5z 2),
One would proceed with the calculations
22pof optimal inventory policies using � � 60 and
� 2 � 650.25 as the mean and the variance of lead time demand.
and the built-in Excel function normsdist( ), which returns the value of �( ). In this
way one could compute (Q, R) policies within the spreadsheet. Alternatively, one
6
In cases where the underlying demand process is nonstationary, that is, where there is trend or
could build a table of L(z), which can then be embedded into a search routine.
seasonality, the variance of demand due to systematic changes could be higher than the variance of
forecast error. All of our analysis assumes stationary (constant mean and variance) demand patterns,
Negative Safety Stock
however.
7
Hadley the
and shortage
Whitin (1963),
153.
When
costp.or
service level is relatively low, it is possible for a negative
safety stock situation to arise. Recall that safety stock is the expected inventory on
hand at the arrival of an order. If the safety stock is negative, there would be a backorder situation in more than 50 percent of the order cycles (that is, the Type 1 service
would be under 50 percent). When this occurs, the optimal z value is negative, and the
optimal L(z) value would exceed L(0) � .3989. We illustrate with an example.
Example 5.8
nah77856_ch05.indd 279
Consider once again Harvey’s Specialty Shop from Examples 5.4 to 5.7. Harvey sells a high-end
espresso machine, an expensive and bulky item. For this reason, Harvey attaches a very high hold17/12/14
7:10 PM
formula into Excel, use the definition of the standard normal density function, which is
1
exp (�0.5z 2),
22p
and the built-in Excel function normsdist( ), which returns the value of �( ). In this
way one could compute (Q, R) policies within the spreadsheet. Alternatively, one
280 Chapter Five Inventory
Control
Subject
to Uncertain
Demand
could
build
a table
of L(z),
which can then be embedded into a search routine.
(z) �
Negative Safety Stock
When the shortage cost or service level is relatively low, it is possible for a negative
safety stock situation to arise. Recall that safety stock is the expected inventory on
hand at the arrival of an order. If the safety stock is negative, there would be a backorder situation in more than 50 percent of the order cycles (that is, the Type 1 service
would be under 50 percent). When this occurs, the optimal z value is negative, and the
optimal L(z) value would exceed L(0) � .3989. We illustrate with an example.
Example 5.8
Consider once again Harvey’s Specialty Shop from Examples 5.4 to 5.7. Harvey sells a high-end
espresso machine, an expensive and bulky item. For this reason, Harvey attaches a very high holding cost of $50 per year to each machine. Harvey sells about 20 of these yearly, and estimates the
variance of yearly demand to be 50. Annual demand follows the normal distribution. Since customers are willing to wait for the machine when he is out of stock, the shortage penalty is low.
He estimates it to be $25 per unit. Order lead time is six months. The fixed cost of ordering is $80.
Assume that the lot size used for reordering is the EOQ value. Determine the following:
1.
2.
3.
4.
Solution
Optimal reorder level
Safety stock
Resulting Type 1 service level
Resulting Type 2 service level
1. We have that K � $80, p � $25, h � $50, and � 20. Scaling lead time demand to
0.5 year, we obtain � 10 and 1(50)(0.5) � 5. The reader can check that the EOQ
rounded to the nearest integer is Q � 8. The equation for determining R0 is
1 � F(R0) �
nah77856_ch05.qxd
1/9/08
(8)(50)
Q0h
�
� 0.8.
p
(25)(20)
Since 1 � F(R0) exceeds 0.5, the resulting z value is negative. In this case z � �0.84,
Page
L(z)
� 279
0.9520, and R0 � z � � (5)(�.84) � 10 � 5.8, which we round to 6. Hence, the
optimal solution based on the EOQ is to order eight units when the on-hand inventory falls to
six units.
2. The safety stock is S � R � � 6 � 10 � �4.
3. The Type 1 service level is � F(R) � .20 (or 20 percent).
4. The Type 2 service level is � 1 � n(R)>Q � 1 � L(z)>Q � 1 (5)(.9520)�8 0.405
5.5 Service Levels in (Q, R) Systems 279
(or 41 percent).
11:42 AM
Note: Had we tried to solve this problem for the optimal (Q, R) iteratively using Equations
(1) and (2), the equations would have diverged and not yielded a solution. This is a result of
having a very low shortage cost compared to the holding cost. Furthermore, if Q were larger,
the problem could be unsolvable. For example, if one assumes the setup cost K 150, then
EOQ 11 and 1 F(R0) Q0hp (11)(50)(25)(20) 1.1, which is obviously not
solvable. Such circumstances are very rare in practice, but, as we see from this example, it is
possible for the model to fail. This is a consequence of the fact that the model is not exact.
Exact (Q, R) models are beyond the scope of this book, and are known only for certain
demand distributions.
Problems for Sections 5.4 and 5.5
nah77856_ch05.indd 280
13. An automotive warehouse stocks a variety of parts that are sold at neighborhood
stores. One particular part, a popular brand of oil filter, is purchased by the
warehouse for $1.50 each. It is estimated that the cost of order processing and
receipt is $100 per order. The company uses an inventory carrying charge based on
a 28 percent annual interest rate.
The monthly demand for the filter follows a normal distribution with mean
280 and standard deviation 77. Order lead time is assumed to be five months.
Assume that if a filter is demanded when the warehouse is out of stock, then the
demand is back-ordered, and the cost assessed for each back-ordered demand is
$12.80. Determine the following quantities:
a. The optimal values of the order quantity and the reorder level.
17/12/14 7:10 PM
Exact (Q, R) models are beyond the scope of this book, and are known only for certain
demand distributions.
Problems for Sections 5.4 and 5.5
13. An automotive warehouse stocks a variety of parts that are sold at neighborhood
stores. One particular part, a popular brand of5.5oilService
filter,
is inpurchased
by the
Levels
(Q, R) Systems 281
warehouse for $1.50 each. It is estimated that the cost of order processing and
receipt is $100 per order. The company uses an inventory carrying charge based on
a 28 percent annual interest rate.
The monthly demand for the filter follows a normal distribution with mean
280 and standard deviation 77. Order lead time is assumed to be five months.
Assume that if a filter is demanded when the warehouse is out of stock, then the
demand is back-ordered, and the cost assessed for each back-ordered demand is
$12.80. Determine the following quantities:
a. The optimal values of the order quantity and the reorder level.
b. The average annual cost of holding, setup, and stock-out associated with this
item assuming that an optimal policy is used.
c. Evaluate the cost of uncertainty for this process. That is, compare the average
annual cost you obtained in part (b) with the average annual cost that would be
incurred if the lead time demand had zero variance.
14. Weiss’s paint store uses a (Q, R) inventory system to control its stock levels. For a
particularly popular white latex paint, historical data show that the distribution of
monthly demand is approximately normal, with mean 28 and standard deviation 8.
Replenishment lead time for this paint is about 14 weeks. Each can of paint costs
the store $6. Although excess demands are back-ordered, the store owner estimates
that unfilled demands cost about $10 each in bookkeeping and loss-of-goodwill
costs. Fixed costs of replenishment are $15 per order, and holding costs are based
on a 30 percent annual rate of interest.
a. What are the optimal lot sizes and reorder points for this brand of paint?
b. What is the optimal safety stock for this paint?
15. After taking a production seminar, Al Weiss, the owner of Weiss’s paint store
mentioned in Problem 14, decides that his stock-out cost of $10 may not be very
accurate and switches to a service level model. He decides to set his lot size by the
EOQ formula and determines his reorder point so that there is no stock-out in
nah77856_ch05.qxd 1/9/08 11:42 AM
Page 280
90 percent
of the order cycles.
a. Find the resulting (Q, R) values.
b. Suppose that, unfortunately, he really wanted to satisfy 90 percent of his
demands (that is, achieve a 90 percent fill rate). What fill rate did he actually
achieve from the policy determined in part (a)?
16.Control
Suppose
in Problem
280 Chapter Five Inventory
Subjectthat
to Uncertain
Demand13 the stock-out cost is replaced with a Type 1 service
objective of 95 percent. Find the optimal values of (Q, R) in this case.
17. Suppose that in Problem 13 a Type 2 service objective of 95 percent is substituted
for the stock-out cost of $12.80. Find the resulting values of Q and R. Also, what
is the imputed cost of shortage for this case?
18. Suppose that the warehouse mistakenly used a Type 1 service objective when it really
meant to use a Type 2 service objective (see Problems 16 and 17). What is the additional holding cost being incurred each year for this item because of this mistake?
19. Disk Drives Limited (DDL) produces a line of internal Winchester disks for
microcomputers. The drives use a 3.5-inch platter that DDL purchases from an
outside supplier. Demand data and sales forecasts indicate that the weekly demand
for the platters seems to be closely approximated by a normal distribution with
mean 38 and variance 130. The platters require a three-week lead time for receipt.
DDL has been using a 40 percent annual interest charge to compute holding costs.
The platters cost $18.80 each, order cost is $75.00 per order, and the company is
currently using a stock-out cost of $400.00 per platter. (Because the industry is so
competitive, stock-outs are very costly.)
a. Because of a prior contractual agreement with the supplier, DDL must purchase
the platters in lots of 500. What is the reorder point that it should be using in this
case?
b. When DDL renegotiates its contract with the supplier, what lot size should it
write into the agreement?
c.
How
much of a penalty in terms of setup, holding, and stock-out cost is DDL
nah77856_ch05.indd 281
17/12/14
7:10 PM
meant to use a Type 2 service objective (see Problems 16 and 17). What is the additional holding cost being incurred each year for this item because of this mistake?
19. Disk Drives Limited (DDL) produces a line of internal Winchester disks for
microcomputers. The drives use a 3.5-inch platter that DDL purchases from an
outside supplier. Demand data and sales forecasts indicate that the weekly demand
for the platters seems to be closely approximated by a normal distribution with
meanSubject
38 and
variance
130. The platters require a three-week lead time for receipt.
282 Chapter Five Inventory Control
to Uncertain
Demand
DDL has been using a 40 percent annual interest charge to compute holding costs.
The platters cost $18.80 each, order cost is $75.00 per order, and the company is
currently using a stock-out cost of $400.00 per platter. (Because the industry is so
competitive, stock-outs are very costly.)
a. Because of a prior contractual agreement with the supplier, DDL must purchase
the platters in lots of 500. What is the reorder point that it should be using in this
case?
b. When DDL renegotiates its contract with the supplier, what lot size should it
write into the agreement?
c. How much of a penalty in terms of setup, holding, and stock-out cost is DDL
paying for contracting to buy too large a lot?
d. DDL’s president is uncomfortable with the $400 stock-out cost and decides to
substitute a 99 percent fill rate criterion. If DDL used a lot size equal to the
EOQ, what would its reorder point be in this case? Also, find the imputed cost
of shortage.
20. Bobbi’s Restaurant in Boise, Idaho, is a popular place for weekend brunch. The
restaurant serves real maple syrup with french toast and pancakes. Bobbi buys
the maple syrup from a company in Maine that requires three weeks for delivery.
The syrup costs Bobbi $4 a bottle and may be purchased in any quantity. Fixed costs
of ordering amount to about $75 for bookkeeping expenses, and holding costs are
based on a 20 percent annual rate. Bobbi estimates that the loss of customer goodwill for not being able to serve the syrup when requested amounts to $25. Based on
past experience, the weekly demand for the syrup is normal with mean 12 and variance 16 (in bottles). For the purposes of your calculations, you may assume that
there are 52 weeks in a year and that all excess demand is back-ordered.
a. How large an order should Bobbi be placing with her supplier for the maple
syrup, and when should those orders be placed?
b. What level of Type 1 service is being provided by the policy you found in part (a)?
c. What level of Type 2 service is being provided by the policy you found in
part (a)?
d. What policy should Bobbi use if the stock-out cost is replaced with a Type 1
nah77856_ch05.qxd 1/9/08 11:42 AM Pageservice
281 objective of 95 percent?
e. What policy should Bobbi use if the stock-out cost is replaced with a Type 2
service objective of 95 percent? (You may assume an EOQ lot size.)
f. Suppose that Bobbi’s supplier requires a minimum order size of 500 bottles.
Find the reorder level that Bobbi should use if she wishes to satisfy 99 percent
of her customer demands for the syrup.
5.6
5.6
Additional Discussion of Periodic-Review Systems 281
ADDITIONAL DISCUSSION OF PERIODIC-REVIEW SYSTEMS
(s, S) Policies
In our analysis of the newsvendor problem, we noted that a severe limitation of the model
from a practical viewpoint is that there is no setup cost included in the formulation.
The (Q, R) model treated in the preceding sections included an order setup cost, but
assumed that the inventory levels were reviewed continuously; that is, known at all
times. How should the system be managed when there is a setup cost for ordering, but
inventory levels are known only at discrete points in time?
The difficulty that arises from trying to implement a continuous-review solution in
a periodic-review environment is that the inventory level is likely to overshoot the
reorder point R during a period, making it impossible to place an order the instant
the inventory reaches R. To overcome this problem, the operating policy is modified
slightly. Define two numbers, s and S, to be used as follows: When the level of on-hand
inventory is less than or equal to s, an order for the difference between the inventory
and S is placed. If u is the starting inventory in any period, then the (s, S) policy is
nah77856_ch05.indd 282
If u � s, order S � u.
If u � s, do not order.
17/12/14 7:10 PM
from a practical viewpoint is that there is no setup cost included in the formulation.
The (Q, R) model treated in the preceding sections included an order setup cost, but
assumed that the inventory levels were reviewed continuously; that is, known at all
times. How should the system be managed when there is a setup cost for ordering, but
inventory levels are known only at discrete points in time?
The difficulty that arises from trying to implement a continuous-review solution in
a periodic-review environment is that the
inventoryDiscussion
level isof likely
to overshoot
the
5.6 Additional
Periodic-Review
Systems 283
reorder point R during a period, making it impossible to place an order the instant
the inventory reaches R. To overcome this problem, the operating policy is modified
slightly. Define two numbers, s and S, to be used as follows: When the level of on-hand
inventory is less than or equal to s, an order for the difference between the inventory
and S is placed. If u is the starting inventory in any period, then the (s, S) policy is
If u � s, order S � u.
If u � s, do not order.
Determining optimal values of (s, S) is extremely difficult, and for that reason few
real operating systems use optimal (s, S) values. Several approximations have been
suggested. One such approximation is to compute a (Q, R) policy using the methods
described earlier, and set s � R and S � R � Q. This approximation will give reasonable results in many cases, and is probably the most commonly used. The reader
interested in a comprehensive comparison of several approximate (s, S) policies should
refer to Porteus (1985).
*Service Levels in Periodic-Review Systems
Service levels also may be used when inventory levels are reviewed periodically. Consider first a Type 1 service objective. That is, we wish to find the order-up-to point Q so
that all the demand is satisfied in a given percentage of the periods. Suppose that the
value of Type 1 service is . Then Q should solve the equation
F(Q) � .
This follows because F(Q) is the probability that the demand during the period does
not exceed Q. Notice that one simply substitutes for the critical ratio in the newsvendor
model. To find Q to satisfy a Type 2 service objective of , it is necessary to obtain an
expression for the fraction of demands that stock out each period. Using essentially the
same notation as that used for (Q, R) systems, define
n(Q) �
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 282
(x � Q)f(x) dx.
�
Q
Note that n(Q), which represents the expected number of demands that stock out at the
end of the period, is the same as the term multiplying cu in the expression for the
expected cost function for the newsvendor model discussed in Section 5.3. As the demand
per period is , it follows that the proportion of demands that stock out each period is
n(Q). Hence, the value of Q that meets a fill rate objective of solves
282 Chapter Five Inventory Control Subject to Uncertain Demand
n(Q) � (1 � ).
The specification of either a Type 1 or Type 2 service objective completely determines the order quantity, independent of the cost parameters.
Example 5.9
Mac, the owner of the newsstand described in Example 5.1, wishes to use a Type 1 service level
of 90 percent to control his replenishment of The Computer Journal. The z value corresponding
to the 90th percentile of the unit normal is z � 1.28. Hence,
Q* � z � � (4.74)(1.28) � 11.73 � 17.8 18.
Using a Type 2 service of 90 percent, we obtain
n(Q) � (1 � ) � (0.1)(11.73) � 1.173.
It follows that L(z) � n(Q) � 1.1734.74 � 0.2475. From Table A–4 at the back of this book,
we find
z 0.35;
then
Q* � z � � (4.74)(0.35) � 11.73 � 13.4 13.
As with (Q, R) models, notice the striking difference between the resulting values of Q* for
the same levels of Type 1 and Type 2 service.
nah77856_ch05.indd 283
17/12/14 7:10 PM
Example 5.9
Mac, the owner of the newsstand described in Example 5.1, wishes to use a Type 1 service level
of 90 percent to control his replenishment of The Computer Journal. The z value corresponding
to the 90th percentile of the unit normal is z � 1.28. Hence,
Q* � z � � (4.74)(1.28) � 11.73 � 17.8 18.
Using a Type 2 service of 90 percent, we obtain
284 Chapter
Five Inventory
256
Chapter Five
Inventory Control Subject to Uncertain Demand
n(Q) � (1 � ) � (0.1)(11.73) � 1.173.
It follows that L(z) � n(Q) � 1.1734.74 � 0.2475. From Table A–4 at the back of this book,
we find
z 0.35;
then
Q* � z � � (4.74)(0.35) � 11.73 � 13.4 13.
As with (Q, R) models, notice the striking difference between the resulting values of Q* for
the same levels of Type 1 and Type 2 service.
Fixed
Problems for Section
5.6Order Size Model
If a positive order lead time is included, only a slight modification of these equations is
21. Consider the Crestview Printing Company mentioned in Problem 9. Suppose
required. In particular, the response time of the system is now the order lead time plus
that Crestview wishes to produce enough cards to satisfy all Christmas demand
one period. Hence, we would now use the distribution of demand over t + T, where
with probability 90 percent. How many cards should they print? Suppose the probT is the time between inventory reviews.
ability is 97 percent. What would you recommend in this case? (Your answer will
This periodic review service level model is very useful in retail settings, in particular.
depend upon the assumption you make concerning the shape of the cumulative
It is common in retailing to place orders at fixed points in time to take advantage of
distribution function.)
bundling multiple orders together.
22. Consider Happy Henry’s car dealer described in Problem 10.
example 5.10
Solution
Problems for
a. How
EX123s
Henry’s
purchaseWisconsin.
to satisfy Stroheim’s
all the demand
Stroheim’s
is amany
dry goods
storeshould
located Happy
in downtown
Milwaukee,
places
orders weekly
their suppliers
for with
all ofprobability
their reorders.
over a with
three-month
interval
.95?The lead time for men’s briefs is
4 days.
a 95
percent
level. if
Assuming
1 service,
is theoforder
b. Stroheim’s
How manyuses
cars
should
be service
purchased
the goala isType
to satisfy
95what
percent
the
up to level for the briefs? Assume demands for briefs are uncertain with daily mean demand
demands?
of 20 and daily standard deviation of 12.
23. For the problem of controlling the inventory of white latex paint at Weiss’s paint
described
in (review
Problemtime
14,plus
suppose
that the
is 11
reordered
a monthly
+ 4 =
The store,
total response
time
lead time)
is 7paint
days. Noonmatter
what
basis of
rather
a continuous
basis.the Central Limit Theorem indicates that the
the form
the than
daily on
demand
distribution,
demand
over 11the
days
be closeyou
to normal.
Thein
parameters
areProblem
m = (11)(20)
= 220 and
a. Using
(Q,should
R) solution
obtained
part (a) of
14, determine
Hence
s = 12211
= 39.80.
appropriate
values
of it(s,follows
S). that for a type 1 service objective of 95 percent, the
order-up-to-point should be Q = sz + m = (39.80)(1.645) + 220 = 286.
b. Suppose that the demands during the months of January to June were
If Stroheim’s was interested in a Type 2 service objective, the z value would be
lower, and the corresponding order up to point would be lower as well. In particular,
and L(z) = n(Q)>sMonth
= 11>39.80 = .2764
, which gives
n(Q) = (1 - Month
b) m = (.05)(220) = 11,
Demand
Demand
a z value of 0.27 and a corresponding order up to level of 231.
January
37
April
31
February
33
May
14
26
June
40
Section 5.6 March
21. Consider
Crestview
Printing
Company
mentioned
in Problem
9. the
Suppose
If thethe
starting
inventory
in January
was 26
cans of paint,
determine
numthatber
Crestview
wishes
to
produce
enough
cards
to
satisfy
all
Christmas
demand
of units of paint ordered in each of the months January to June following
the
with
90 found
percent.
many cards should they print? Suppose the prob(s,probability
S) policy you
in How
part (a).
ability is 97 percent. What would you recommend in this case? (Your answer will
depend upon the assumption you make concerning the shape of the cumulative
distribution function.)
22. Consider Happy Henry’s car dealer described in Problem 10.
a. How many EX123s should Happy Henry’s purchase to satisfy all the demand
over a three-month interval with probability .95?
b. How many cars should be purchased if the goal is to satisfy 95 percent of the
demands?
284
nah77856_ch05.indd 256
17/12/14 6:18
7:10 PM
17/11/14
/14 6:18 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 283
Snapshot Application
TROPICANA USES SOPHISTICATED MODELING
FOR INVENTORY MANAGEMENT
Tropicana, based in Bradenton, Florida, is one of the
world’s largest suppliers of citrus-based juice products.
The company was founded by an Italian immigrant,
Anthony Rossi, in 1947, and was acquired by PepsiCo in
1998. From its production facilities in Florida, Tropicana
makes daily rail shipments to its regional distribution
centers (DCs).
The focus of this application is the largest DC located
in Jersey City, New Jersey, that services the northeast
United States and Canada. Shipments from Bradenton
to Jersey City require four days: one day for loading,
two days in transit, and one day for unloading. This is
the order lead time as seen from the DC. Lead time variability is not considered to be a significant issue.
Based on a statistical analysis of past data, Tropicana
planners have determined that daily demands on the
DC are closely approximated by a normal distribution
for each of their products. Product classes are assumed
to be independent.
Trains are sent from Florida five times per week;
arrivals coincide with each business day at the Jersey
City DC. Demand data and inventory levels are reviewed very frequently, a reorder point R triggers replenishment, and lot size Q is defined by the user.
5.7
5.6 Additional Discussion of Periodic-Review Systems 285
Hence, they operate a standard (Q, R) continuousreview system with a positive lead time. The state variable
is the inventory position defined as the total amount of
stock on hand at, and in transit to, the DC.
Let D and D be the mean and standard deviation of
daily demand for a particular product line. According to
the theory outlined in this chapter, the mean and standard deviation of demand over the four-day lead time
should be
D 4D,
D 1 2D.
However, analysis of Tropicana data showed that the
standard deviation of lead time demand is closer to
D0.7 2.64D.
The firm’s objective is to maintain a 99.5 percent
Type 2 service level at the DC. Values of (Q, R) are computed using the methodology of this chapter. Planners
check the inventory on hand and in transit, and place
an order equal to the EOQ, when this value falls below
the reorder level, R. This analysis is carried out for a
wide range of the company’s products and individually
determined for each regional DC.
Source: Based on joint work between the author and Tim
Rowell of Tropicana.
MULTIPRODUCT SYSTEMS
23. For the problem of controlling the inventory of white latex paint at Weiss’s paint
ABC Analysis
store, described in Problem 14, suppose that the paint is reordered on a monthly
One basis
issuerather
that we
have
discussed is
the cost of implementing an inventory control
than
onnot
a continuous
basis.
system and the trade-offs between the cost of controlling the system and the potential
a. Using the (Q, R) solution you obtained in part (a) of Problem 14, determine
benefits that accrue from that control. In multiproduct inventory systems, not all
­appropriate values of (s, S).
products are equally profitable. Control costs may be justified in some cases and not in
b.
that the
demands
during
the months
of January
June
were
others. Suppose
For example,
spending
$200
annually
to monitor
an itemtothat
contributes
only
$100 a year to profits is clearly not economical.
For this
reason, it is importantDemand
to differentiate profitable
from unprofitable
items. To
Month
Month
Demand
do so, we borrow a concept from economics. The economist Vilfredo Pareto, who stud37 19th century, noted
April that a large portion
31
ied theJanuary
distribution of wealth in the
of the
33 of the population.
MayThis Pareto effect14also applies
wealth February
was owned by a small segment
March systems: a large portion
26 of the total dollar
June volume of sales
40 is often
to inventory
accounted for by a small number of inventory items. Assume that items are ranked in
the starting
in January
was 26
cansThe
of paint,
determine
the of
­number
decreasingIforder
of theinventory
dollar value
of annual
sales.
cumulative
value
sales
of results
units ofinpaint
ordered
each
thepictured
monthsinJanuary
June ­following the
generally
a curve
muchinlike
theofone
Figure to
5–8.
(s, S) policy
found inofpart
Typically,
the topyou
20 percent
the(a).
items account for about 80 percent of the annual
dollar volume of sales, the next 30 percent of the items for the next 15 percent of sales,
and the remaining 50 percent for the last 5 percent of dollar volume. These figures are
nah77856_ch05.indd 285
17/12/14 7:10 PM
the reorder level, R. This analysis is carried out for a
wide range of the company’s products and individually
determined for each regional DC.
Trains are sent from Florida five times per week;
arrivals coincide with each business day at the Jersey
City DC. Demand data and inventory levels are reviewed very frequently, a reorder point R triggers replenishment, and lot size Q is defined by the user.
Source: Based on joint work between the author and Tim
Rowell of Tropicana.
286 Chapter Five Inventory Control Subject to Uncertain Demand
5.7
MULTIPRODUCT SYSTEMS
ABC Analysis
nah77856_ch05.qxd
284
1/9/08
One issue that we have not discussed is the cost of implementing an inventory control
system and the trade-offs between the cost of controlling the system and the potential
benefits that accrue from that control. In multiproduct inventory systems, not all
products are equally profitable. Control costs may be justified in some cases and not in
others. For example, spending $200 annually to monitor an item that contributes only
$100 a year to profits is clearly not economical.
For this reason, it is important to differentiate profitable from unprofitable items. To
do so, we borrow a concept from economics. The economist Vilfredo Pareto, who studied the distribution of wealth in the 19th century, noted that a large portion of the
wealth was owned by a small segment of the population. This Pareto effect also applies
to inventory systems: a large portion of the total dollar volume of sales is often
11:42
AM Page
284a small number of inventory items. Assume that items are ranked in
accounted
for by
decreasing order of the dollar value of annual sales. The cumulative value of sales
generally results in a curve much like the one pictured in Figure 5–8.
Typically, the top 20 percent of the items account for about 80 percent of the annual
dollar volume of sales, the next 30 percent of the items for the next 15 percent of sales,
and the remaining 50 percent for the last 5 percent of dollar volume. These figures are
Chapter Five Inventory Control Subject to Uncertain Demand
FIGURE 5–8
Pareto curve:
Distribution of
inventory by value
95%
80%
1
0.9
0.8
Cumulative fraction of
value of inventory
0.7
0.6
0.5
0.4
0.3
0.2
A
0.1
0
0
B
C
0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90
1
Proportion of inventory items
20%
nah77856_ch05.indd 286
50%
only approximate and will vary slightly from one system to another. The three item
groups are labeled A, B, and C, respectively. When a finer distinction is needed, four or
five categories could be used. Even when using only three categories, the percentages
used in defining A, B, and C items could be different from the 80 percent, 15 percent, and
5 percent recommended.
Because A items account for the lion’s share of the yearly revenue, these items
should be watched most closely. Inventory levels for A items should be monitored
continuously. More sophisticated forecasting procedures might be used and more care
would be taken in the estimation of the various cost parameters required in calculating
operating policies. For B items inventories could be reviewed periodically, items could
be ordered in groups rather than individually, and somewhat less sophisticated forecasting methods could be used. The minimum degree of control would be applied to
17/12/14 7:10 PM
only approximate and will vary slightly from one system to another. The three item
groups are labeled A, B, and C, respectively. When a finer distinction is needed, four or
five categories could be used. Even when using only three categories, the percentages
used in defining A, B, and C items could be different from the 80
percent, 15 percent,
and
5.7 Multiproduct
Systems 287
5 percent recommended.
Because A items account for the lion’s share of the yearly revenue, these items
should be watched most closely. Inventory levels for A items should be monitored
continuously. More sophisticated forecasting procedures might be used and more care
would be taken in the estimation of the various cost parameters required in calculating
operating policies. For B items inventories could be reviewed periodically, items could
be ordered in groups rather than individually, and somewhat less sophisticated forecasting methods could be used. The minimum degree of control would be applied to
C items. For very inexpensive C items with moderate levels of demand, large lot sizes
are recommended to minimize the frequency that these items are ordered. For
expensive C items with very low demand, the best policy is generally not to hold any
inventory. One would simply order these items as they are demanded.
Example 5.10
(continued)
nah77856_ch05.qxd
1/9/08
A sample of 20 different stock items from Harvey’s Specialty Shop is selected at random. These
items vary in price from $0.25 to $24.99 and in average yearly demand from 12 to 786. The
results of the sampling are presented in Table 5–1. In Table 5–2 the items are ranked in decreasing order of the annual dollar volume of sales. Notice that only 4 of the 20 stock items account
for over 80 percent of the annual dollar volume generated by the entire group. Also notice that
11:42 AMthere
Page
are 285
high-priced items in both categories A and C.
This report was very illuminating to Harvey, who had assumed that R077, a packaged goat
cheese from the south of France, was a profitable item because of its cost, and had been
virtually ignoring TTR77, a domestic chocolate bar.
5.7
TABLE 5–1
Performance of 20
Stock Items Selected
at Random
TABLE 5–2
Twenty Stock Items
Ranked in
Decreasing Order of
Annual Dollar
Volume
nah77856_ch05.indd 287
Part Number
Price
Yearly Demand
5497J
3K62
88450
P001
2M993
4040
W76
JJ335
R077
70779
4J65E
334Y
8ST4
16113
45000
7878
6193L
TTR77
39SS5
93939
$2.25
2.85
1.50
0.77
4.45
6.10
3.10
1.32
12.80
24.99
7.75
0.68
0.25
3.89
7.70
6.22
0.85
0.77
1.23
4.05
260
43
21
388
612
220
110
786
14
334
24
77
56
89
675
66
148
690
52
12
Dollar Volume
$
Part
Number
Price
Yearly
Demand
Dollar
Volume
70779
45000
2M993
4040
$24.99
7.70
4.45
6.10
334
675
612
220
$8,346.66
5,197.50
2,723.40
1,342.00
$ 8,346.66
13,544.16
16,267.56
17,609.56
1.32
2.25
0.77
786
260
690
1,037.52
585.00
531.30
18,647.08
19,232.08
19,763.38
JJ335
5497J
TTR77
Multiproduct Systems 285
585.00
122.55
31.50
298.76
2,723.40
1,342.00
341.00
1,037.52
179.20
8,346.66
186.00
52.36
14.00
346.21
5,197.50
410.52
125.80
531.30
63.96
48.60
Cumulative
Dollar Volume
A items:
20% of items
account for 80.1%
of total value.
B items:
30% of items
17/12/14 7:10 PM
16113
45000
7878
6193L
TTR77
39SS5
93939
3.89
7.70
6.22
0.85
0.77
1.23
4.05
89
675
66
148
690
52
12
346.21
5,197.50
410.52
125.80
531.30
63.96
48.60
288 Chapter Five Inventory Control Subject to Uncertain Demand
TABLE 5–2
Twenty Stock Items
Ranked in
Decreasing Order of
Annual Dollar
Volume
Part
Number
Price
Yearly
Demand
Dollar
Volume
70779
45000
2M993
4040
$24.99
7.70
4.45
6.10
334
675
612
220
$8,346.66
5,197.50
2,723.40
1,342.00
$ 8,346.66
13,544.16
16,267.56
17,609.56
A items:
20% of items
account for 80.1%
of total value.
1.32
2.25
0.77
6.22
3.89
3.10
786
260
690
66
89
110
1,037.52
585.00
531.30
410.52
346.21
341.00
18,647.08
19,232.08
19,763.38
20,173.90
20,520.11
20,861.11
B items:
30% of items
account for 14.8%
of total value.
P001
0.77
4J65E
7.75
R077
12.80
6193L
0.85
3K62
2.85
39SS5
1.23
334Y
0.68
93939
4.05
88450
1.50
8ST4
0.25
11:50
AM Page 286
388
24
14
148
43
52
77
12
21
56
298.76
186.00
179.20
125.80
122.55
63.96
52.36
48.60
31.50
14.00
21,159.87
21,345.87
21,525.07
21,650.87
21,773.42
21,837.38
21,889.74
21,938.34
21,969.84
21,983.84
JJ335
5497J
TTR77
7878
16113
W76
nah77856_ch05.qxd
1/9/08
Cumulative
Dollar Volume
C items:
50% of items
account for 5.1%
of total value.
Exchange Curves
286
Much of our analysis assumes a single item in isolation and that the relevant cost paramh in the case of service levels) are constants with “correct” values that can be determined. However, it may be more appropriate to think of
one or all of the cost parameters as policy variables. The correct values are those that
result in a control system with characteristics that meet the needs of the firm and the
goals of management. In a typical multiproduct system, the same values of setup cost K
and interest rate I are used for all items. We can treat the ratio KI as a policy variable;
if this ratio is large, lot sizes will be larger and the average investment in inventory will
be greater. If this ratio is small, the number of annual replenishments will increase.
To see exactly how a typical exchange curve is derived, consider a deterministic
system consisting of n products with varying demand rates 1, . . . , n and item values
c1, . . . , cn. If EOQ values are used to replenish stock for each item, then
Chapter Five Inventory
Control
Subject
eters
K, h,
and top Uncertain
(or just Demand
K and
Qi �
nah77856_ch05.indd 288
A
2Ki
Ici
for 1 … i … n.
For item i, the cycle time is Qii, so that iQi is the number of replenishments in
one year. The total number of replenishments for the entire system is iQi. The
average on-hand inventory of item i is Qi2, and the value of this inventory in dollars
is ciQi2. Hence, the total value of the inventory is ciQi2.
Each choice of the ratio KI will result in a different value of the number of replenishments per year and the dollar value of inventory. As KI is varied, one traces out a
curve such as the one pictured in Figure 5–9. An exchange curve such as this one
allows management to easily see the trade-off between the dollar investment in inventory and the frequency of stock replenishment.
Exchange curves also can be used to compare various safety stock and service level
strategies. As an example, consider a system in which a fill rate constraint is used
(i.e., Type 2 service) for all items. Furthermore, suppose that the lead time demand distribution for all items is normal, and each item gets equal service. The dollar value of
the safety stock is ci(Ri � i), and the annual value of back-ordered demand is
ci in(Ri)Qi. A fixed value of the fill rate will result in a set of values of the control
variables (Q1, R1), . . . , (Qn, Rn), which can be computed by the methods discussed
17/12/14 7:10 PM
For item i, the cycle time is Qii, so that iQi is the number of replenishments in
one year. The total number of replenishments for the entire system is iQi. The
average on-hand inventory of item i is Qi2, and the value of this inventory in dollars
is ciQi2. Hence, the total value of the inventory is ciQi2.
Each choice of the ratio KI will result in a different value of the number of replenishments per year and the dollar value of inventory. As KI is varied, one traces out a
curve such as the one pictured in Figure 5–9. An exchange curve such as this one
5.7 Multiproduct
Systems 289
allows management to easily see the trade-off between the dollar
investment
in inventory and the frequency of stock replenishment.
Exchange curves also can be used to compare various safety stock and service level
strategies. As an example, consider a system in which a fill rate constraint is used
(i.e., Type 2 service) for all items. Furthermore, suppose that the lead time demand distribution for all items is normal, and each item gets equal service. The dollar value of
the safety stock is ci(Ri � i), and the annual value of back-ordered demand is
ci in(Ri)Qi. A fixed value of the fill rate will result in a set of values of the control
variables (Q1, R1), . . . , (Qn, Rn), which can be computed by the methods discussed
FIGURE 5–9
Exchange curve of
replenishment
frequency and
inventory value
$ value
of average
inventory
n
i =1
ciQi
2
Increasing Kl
Replenishment per year
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 287
n
i =1
5.7
i
Qi
Multiproduct Systems 287
Exchange curve of the
investment in safety
stock and Safety stock ($)
FIGURE 5–10
Increasing Expected number of stock-outs per year
earlier in the chapter. Each set of (Q, R) values yields a pair of values for the safety
stock and the back-ordered demand. As the fill rate is increased, the investment in
safety stock increases and the value of back-ordered demand decreases. The exchange
curve one would obtain is pictured in Figure 5–10. Such an exchange curve is a useful
way for management to assess the dollar impact of various service levels.
Example 5.11
nah77856_ch05.indd 289
Consider the 20 stock items listed in Tables 5–1 and 5–2. Suppose that Harvey, the owner
17/12/14
7:10 PM
Expected number of stock-outs per year
290 Chapter Five Inventory Control Subject to Uncertain Demand
earlier in the chapter. Each set of (Q, R) values yields a pair of values for the safety
stock and the back-ordered demand. As the fill rate is increased, the investment in
safety stock increases and the value of back-ordered demand decreases. The exchange
curve one would obtain is pictured in Figure 5–10. Such an exchange curve is a useful
way for management to assess the dollar impact of various service levels.
Example 5.11
Consider the 20 stock items listed in Tables 5–1 and 5–2. Suppose that Harvey, the owner
of Harvey’s Specialty Shop, is reconsidering his choices of the setup cost of $50 and interest
charge of 20 percent. Harvey uses the EOQ formula to compute lot sizes for the 20 items for a
range of values of KI from 50 to 500. The resulting exchange curve appears in Figure 5–11.
FIGURE 5–11
8
$ Value of average
inventory (thousands)
Exchange curve:
Harvey’s Specialty
Shop
500
450
400
350
300
7
6
250
200
5
150
4
100
3
nah77856_ch05.qxd
1/9/08
11:42 AM
50
Page 288
2
14
18
22
26
30
34
38
42
48
Number of orders per year
50
288 Chapter Five Inventory Control Subject to Uncertain Demand
Harvey is currently operating at KI 500.2 250, which results in approximately 22
orders per year and an average inventory cost of $5,447 annually. By reducing KI to 100, the
inventory cost for these 20 items is reduced to $3,445 and the order frequency is increased to
34 orders a year. After some thought, Harvey decides that the additional time and bookkeeping expenses required to track an additional 12 orders annually is definitely worth the $2,000
savings in inventory cost. (He is fairly comfortable with the 20 percent interest rate, which
means that the true value of his setup cost for ordering is closer to $20 than $50. In this way
the exchange curve can assist in determining the correct value of cost parameters that may
otherwise be difficult to estimate.) He also considers moving to KI 50, but decides that the
additional savings of about $1,000 are not worth having to process almost 50 orders a year.
Problems for Section 5.7
24. Describe the ABC classification system. What is the purpose of classifying items
in this fashion? What would be the primary value of ABC analysis to a retailer? To
a manufacturer?
25. Consider the following list of retail items sold in a small neighborhood gift shop.
Item
nah77856_ch05.indd 290
Greeting cards
T-shirts
Men’s jewelry
Novelty gifts
Children’s clothes
Chocolate cookies
Annual Volume
Average Profit
per Item
3,870
1,550
875
2,050
575
7,000
$ 0.40
1.25
4.50
12.25
6.85
0.10
17/12/14 7:10 PM
otherwise be difficult to estimate.) He also considers moving to KI 50, but decides that the
additional savings of about $1,000 are not worth having to process almost 50 orders a year.
Problems for Section 5.7
24. Describe the ABC classification system. What is the purpose of classifying items
in this fashion? What would be the primary value of5.8 Overview
ABC analysis
to a retailer?
To
of Advanced
Topics 291
a manufacturer?
25. Consider the following list of retail items sold in a small neighborhood gift shop.
Item
Annual Volume
Average Profit
per Item
3,870
1,550
875
2,050
575
7,000
1,285
1,900
$ 0.40
1.25
4.50
12.25
6.85
0.10
3.50
15.00
Greeting cards
T-shirts
Men’s jewelry
Novelty gifts
Children’s clothes
Chocolate cookies
Earrings
Other costume jewelry
a. Rank the item categories in decreasing order of the annual profit. Classify each
in one of the categories as A, B, or C.
b. For what reason might the store proprietor choose to sell the chocolate cookies
even though they might be her least profitable item?
26. From management’s point of view, what is the primary value of an exchange curve?
Discuss both the exchange curve for replenishment frequency and inventory value
and the exchange curve for expected number of stock-outs per year and the investment in safety stock.
27. Consider the eight stock items listed in Problem 25. Suppose that the average costs
of these item categories are
Item
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 289
Greeting cards
T-shirts
Men’s jewelry
Novelty gifts
Children’s clothes
Chocolate cookies
Earrings
Other costume jewelry
Cost
$ 0.50
3.00
8.00
12.50
8.80
0.40
4.80
12.00
5.8
Overview of Advanced Topics 289
Compare the total number of replenishments per year and the dollar value of the
inventory of these items for the following values of the ratio of KI: 100, 200, 500,
1,000. From the four points obtained, estimate the exchange curve of replenishment
frequency and inventory value.
*5.8
OVERVIEW OF ADVANCED TOPICS
This chapter has treated only a small portion of inventory models available. Considerable research has been devoted to analyzing far more complex stochastic inventory
control problems, but most of this research is beyond the scope of our coverage. This
section presents a brief overview of two areas not discussed in this chapter that account
for a large portion of the recent research on inventory management.
Multi-echelon Systems
nah77856_ch05.indd 291
Firms involved in the manufacture and distribution of consumer products must take
into account the interactions of the various levels in the distribution chain. Typically,
items are produced at one or more factories, shipped to warehouses for intermediate
storage, and subsequently shipped to retail outlets to be sold to the consumer. We refer
to the levels of the system as echelons. For example, the factory–warehouse–retailer
three-echelon system is pictured in Figure 5–12.
A system of the type pictured in Figure 5–12 is generally referred to as a distribution system. In such a system, the demand arises at the lowest echelon (the retailer in
17/12/14
7:10 PM
which are eventually combined to form end items (that is, final products). A typical
assembly system is pictured in Figure 5–13.
In the assembly system, external demand originates at the end-item level only.
Demand for the components arises only when “orders” are placed at higher levels of
the system, which are the consequence of end-item production schedules. Materials
requirements planning methodology is designed to model systems of this type, but
5.8 Overview of Advanced Topics 289
does not consider the effect of uncertainty on the final demand. Although there has
292 Chapter Five Inventory
Control
Subject
to Uncertain
Demand
been
some
research
on this
problem, precisely how the uncertainty in the final demand
Compare the total number of replenishments per year and the dollar value of the
inventory of these items for the following values of the ratio of KI: 100, 200, 500,
1,000. From the four points obtained,
estimate the exchange curve of replenishment
Factory
FIGURE 5–12
frequency and inventory value.
Typical three-level
distribution system
*5.8
OVERVIEW OF ADVANCED TOPICS
Warehouse 1
Warehouse 2
Demand
This chapter has treated only a small portion of inventory models available.
Considerable research has been devoted to analyzing far more complex stochastic inventory
control problems, but most of this research is beyond the scope of our coverage. This
section presents a brief overview of two areas not discussed in this chapter that account
for a large portion
of the recent research
on inventory management.
Retailer 1
Retailer 2
Retailer 3
Multi-echelon Systems
Firms involved in the manufacture and distribution of consumer products must take
into account the interactions of the various levels in the distribution chain. Typically,
items are produced at one or more factories, shipped to warehouses for intermediate
storage, and subsequently shipped to retail outlets to be sold to the consumer. We refer
to the levels of the system as echelons. For example, the factory–warehouse–retailer
three-echelon system is pictured in Figure 5–12.
A system of the type pictured in Figure 5–12 is generally referred to as a distribution system. In such a system, the demand arises at the lowest echelon (the retailer in
this case) and is transmitted up to the higher echelons. Production plans at the factory
must be coordinated with orders placed at the warehouse by the retailers. Another type
of multi-echelon system arising in manufacturing contexts is known as an assembly
system. In an assembly system, components are combined to form subassemblies,
which are eventually combined to form end items (that is, final products). A typical
nah77856_ch05.qxd 1/9/08 11:42
AM Page
290is pictured in Figure 5–13.
assembly
system
In the assembly system, external demand originates at the end-item level only.
Demand for the components arises only when “orders” are placed at higher levels of
the system, which are the consequence of end-item production schedules. Materials
requirements planning methodology is designed to model systems of this type, but
does not consider the effect of uncertainty on the final demand. Although there has
290 Chapter Five Inventory
Control
Subject
to Uncertain
Demand
been
some
research
on this
problem, precisely how the uncertainty in the final demand
FIGURE 5–13
Typical
three-level
FIGURE
5–12
R1
assembly
system
Typical three-level
distribution system
Factory
C1
Warehouse 1
R2
Warehouse 2
FP
C2
Retailer 1
R3
Retailer 2
Demand
Demand
Final product
Retailer 3
Components
Raw materials
nah77856_ch05.indd 292
affects the optimal production and replenishment policies of components remains an
open question.
One multi-echelon system that has received a great deal of attention in the literature
(Sherbrooke, 1968) was designed for an application in military logistics. Consider the
replenishment problem that arises in maintaining a supply of spare parts to support jet
17/12/14 7:10 PM
5.8 Overview of Advanced Topics 293
affects the optimal production and replenishment policies of components remains an
open question.
One multi-echelon system that has received a great deal of attention in the literature (Sherbrooke, 1968) was designed for an application in military logistics.
Consider the replenishment problem that arises in maintaining a supply of spare
parts to support jet aircraft engine repair. When an engine (or engine component)
fails, it is sent to a central depot for repair. The depot maintains its own inventory
of engines, so that it can ship a replacement to the base before the repair is completed. The problem is to determine the correct levels of spares at the base and at
the depot given the conflicting goals of budget limitations and the desirability of a
high system fill rate.
The U.S. military’s investment in spare parts is enormous. As far back as 1968
it was estimated that the U.S. Air Force alone has $10 billion invested in repairable item inventories. Considering the effects of inflation and the sizes of the other
armed ­services, the current investment in repairable items in the U.S. military probably ­exceeds $100 billion. Mathematical inventory models have made a significant
impact on the management of ­repairable item inventories, both in military and nonmilitary ­environments.
In the private sector, retailing is an area in which multi-echelon distribution ­systems are common. Most major retail chains utilize a distribution center
(DC) as an ­intermediate storage point between the manufacturer and the retail
store. ­Determining how to allocate inventory between the DC and the stores is an
­important strategic issue. DC inventory allows “risk pooling” among the stores and
facilitates redistribution of store inventories that might grow out of balance. Several
­commercial computer-based inventory packages are available for retail inventory
management, perhaps the best known being Inforem, designed and marketed by
IBM. Readers interested in a ­comprehensive review of the research on the retailer
inventory management problem should refer to the ­recent book by Agrawal and
Smith (2008) and Nahmias and Smith (1992).
Perishable Inventory Problems
Although certain types of decay can be accounted for by adjustments to the holding
cost, the mathematical models that describe perishability, decay, or obsolescence are
generally quite complex.
Typical examples of fixed-life perishable inventories include food, blood, drugs,
and photographic film. In order to keep track of the age of units in stock, it is necessary to know the amount of each age level of inventory on hand. This results in a
complicated multidimensional system state space. Perishability adds an additional
penalty for holding inventory, thus reducing the size of orders placed. The most
notable a­ pplication of the theory of perishable inventory models has been to the
management of blood banks.
A somewhat simpler but related problem is that of exponential decay, in which it is
assumed that a fixed fraction of the inventory in stock is lost each period. Exponential
decay is an accurate description of the changes that take place in volatile liquids such
as alcohol and gasoline, and is an exact model of decline in radioactivity of ­radioactive
pharmaceuticals and nuclear fuels.
Obsolescence has been modeled by assuming that the length of the planning ­horizon
is a random variable with a known probability distribution. In practice, such ­models
nah77856_ch05.indd 293
17/12/14 7:10 PM
Snapshot Application
294 Chapter Five Inventory Control Subject to Uncertain Demand
Intel uses Multiechelon Inventory
­Modelling to Manage the Supply Chain
for Boxed CPUs
Intel Corporation, headquartered in Santa Clara,
California, is the largest producer of integrated
­
­circuits in the world. Intel microprocessors have been
the “guts” of most personal computers ever since IBM
first adopted the Intel 8088 as the CPU in the original
IBM PC in 1981. In 2005, a group at Intel and one
outside consultant was charged with the problem
of managing the supply chain of its branded boxed
CPUs. The division typically sells 100–150 configurations of these products.
The group chose to model the problem as a three
echelon inventory system: (1) global supply, (2) boxing
sites, and (3) boxed CPU warehouses. The group quickly
realized that there were many factors affecting the supply chain, but the primary driver of system effectiveness
was demand uncertainty. Predicting demand uncertainty is difficult in this environment since product life
cycles are often relatively short, resulting in very little
historical data for some products.
Another complicating factor is that the three echelons are far apart geographically, resulting in a very
long cumulative lead time for the system. Hence, a
make-to-order business model is simply not feasible. For
that reason, proper positioning of inventory along the
entire supply chain is critical. The mathematical model
developed for this application was a nonlinear mathematical programming model that could be solved by
standard dynamic programming techniques.
A key issue is metrics. That is, how does one evaluate the quality of the model’s results? One metric is the
total inventory in the system. This translates to the total
dollar investment required by Intel. A second metric is
the quality of the service provided to the customer. Both
Type 1 and Type 2 service level metrics (as discussed in
detail in this chapter) were considered. Intel management chose a modified version of Type 2 service (that
is, the fill rate). Their metric was the percentage of an
order filled on the date the customer asks for d
­ elivery
of the product.
The team decided that it was not worth trying to apply
their optimization tool to all 150 products, so they embarked on an ABC analysis, and ultimately chose to apply
the model only to products with more than 3,000 sales per
month and 12,000 sales per quarter. Also, it became clear
that in order to be responsive to customer demand, the
vast majority of the inventory would be held at the third
echelon, namely the boxed CPU warehouses. This inventory has the most value added, so is the most expensive.
That expense is offset by the fact that keeping the inventory closer to the customer significantly improves service
levels. The system was implemented at Intel in late 2005.
Comparisons with pre and post implementation metrics
showed that the model ­resulted in a modest decline in
system-wide inventory levels, but significantly improved
level of customer service.
An important aspect of this study was the priorities
set by the project team. These priorities provide a good
blueprint on how to go about a project of this type to
maximize the likelihood of implementation. The four
goals they set were: (1) keep models and processes simple,
(2) make things better now, (3) implement changes in a
phased manner, and (4) be clear about what success is.
Source: Wieland, B. et al. “Optimizing Inventory Levels within
Intel’s Channel Supply Demand Operations.” Interfaces 42
(2012), pp. 517–527.
are valuable only if the distribution of the useful lifetime of the item can be ­accurately
­estimated in advance.
A related problem is the management of style goods, such as fashion items
in the garment industry. The style goods problem differs from the inventory
models considered in this chapter in that not only is the demand itself uncertain,
but the distribution of demand is uncertain as well. This is typical of problems
for which there is no prior demand history. The procedures suggested to deal
with the style goods problem generally involve some type of Bayesian updating
scheme to combine current observations of demand with prior estimates of the
demand distribution.
nah77856_ch05.indd 294
17/12/14 7:10 PM
nah77856_ch05.qxd
292
5.9
1/9/08
11:42 AM
Page 292
Chapter Five Inventory Control Subject to Uncertain Demand
5.9 Historical Notes and Additional Readings 295
HISTORICAL NOTES AND ADDITIONAL READINGS
Research on stochastic inventory models was sparked by the national war effort and
appears to date back to the 1940s, although the first published material appeared in the
early 1950s (see Arrow, Harris, and Marschak, 1951; and Dvorestsky, Kiefer, and
Wolfowitz, 1952). The important text by Arrow, Karlin, and Scarf (1958) produced
considerable interest in the field.
The model we discuss of a lot size–reorder point system with stock-out cost criterion seems to be attributed to Whitin (1957). The extensions to service levels are
attributed to Brown (1967). Brown also treats a variety of other topics of practical
importance, including the relationship between forecasting models and inventory control. An excellent discussion of exchange curves and ABC analysis, and more in-depth
coverage of a number of the topics we discuss, can be found in Silver and Peterson
(1985). Love (1979) provides an excellent summary of many of the issues we treat.
Hadley and Whitin (1963) also deal with these topics at a more sophisticated mathematical level.
There has been considerable interest in the development of approximate (s, S )
policies. Scarf (1960), and later Iglehart (1963), proved the optimality of (s, S) policies,
and Veinott and Wagner (1965) considered methods for computing optimal (s, S)
policies. A number of approximation techniques have been suggested. Ehrhardt (1979)
considers using regression analysis to fit a grid of optimal policies, whereas Freeland
and Porteus (1980) adapt dynamic programming methods to the problem. Porteus
(1985) numerically compares the effectiveness of various approximation techniques.
Some of the issues regarding lead time uncertainty are treated by Kaplan (1970) and
Nahmias (1979), among others.
The original formulation of the standard multi-echelon inventory problem came
from Clark and Scarf (1960). Extensions have been developed by a variety of researchers, most notably Bessler and Veinott (1966), and, more recently, Federgruen and
Zipkin (1984). Retailer-warehouse systems have been studied by Deuermeyer and
Schwarz (1981) and Eppen and Schrage (1981). Schmidt and Nahmias (1985) analyzed
an assembly system when demand for the final product is random.
The classic work on the type of multi-echelon model that has become the basis for
the inventory control systems implemented in the military was done by Sherbrooke
(1968). An extension of Sherbrooke’s analysis is considered by Graves (1985).
Muckstadt and Thomas (1980) discuss the advantages of implementing multi-echelon
models in an industrial environment. A comprehensive review of models for managing
repairables can be found in Nahmias (1981).
Interest in perishable inventory control models appears to stem from the problem of
blood bank inventory control, although food management has a far greater economic
impact. Most of the mathematical models for perishables assume that the inventory is
issued from stock on an oldest-first basis (although a notable exception is the study by
Cohen and Pekelman, 1978). Nahmias (1982) provides a comprehensive review of the
research on perishable inventory control.
The style goods problem has been studied by a variety of researchers. Representative work in this area includes that by Murray and Silver (1966), Hartung (1973), and
Hausman and Peterson (1972). Those interested in additional readings on stochastic
inventory models should refer to the various review articles in the field. Some of the
most notable include those of Scarf (1963), Veinott (1966), and Nahmias (1978). A
recent collection of up-to-date reviews can be found in Graves et al. (1992).
nah77856_ch05.indd 295
17/12/14 7:10 PM
nah77856_ch05.qxd
1/17/08
3:53 PM
Page 293
296 Chapter Five Inventory Control Subject to Uncertain Demand
5.10 Summary
5.10
Summary 293
This chapter presented an overview of several inventory control methods when the demand
for the item is random. The newsvendor model is based on the assumption that the product has
a useful life of exactly one planning period. We assumed that the demand during the period
is a continuous random variable with cumulative distribution function F(x), and that there
are specified overage and underage costs of co and cu charged against the inventory remaining on hand at the end of the period or the excess demand, respectively. The optimal order
quantity Q* solves the equation
F(Q*) �
cu
.
cu � co
Extensions to discrete demand and multiple planning periods also were considered.
From a practical standpoint, the newsvendor model has a serious limitation: it does not
allow for a positive order setup cost. For that reason, we considered an extension of the
EOQ model known as a lot size–reorder point model. The key random variable for this case
was the demand during the lead time. We showed how optimal values of the decision
variables Q and R could be obtained by the iterative solution of two equations. The system
operates in the following manner. When the level of on-hand inventory hits R, an order for
Q units is placed (which will arrive after the lead time ). This policy is the basis of many
commercial inventory control systems.
Service levels provide an alternative to stock-out costs. Two service levels were
considered: Type 1 and Type 2. Type 1 service is the probability of not stocking out in any
order cycle, and Type 2 service is the probability of being able to meet a demand when it
occurs. Type 2 service, also known as the fill rate, is the more natural definition of service for
most applications.
Several additional topics for (Q, R) systems also were considered. The imputed
shortage cost is the effective shortage cost resulting from specification of the service level. Assuming normality, we showed how one transforms the distribution of
periodic demand to lead time demand. Finally, we considered the effects of lead time
variability.
If a setup cost is included in the multiperiod version of the newsvendor problem, the optimal form of the control policy is known as an (s, S) policy. This means that if the starting
inventory in a period, u, is less than or equal to s, an order for S � u units is placed. An
effective approximation for the optimal (s, S) policy can be obtained by solving the
problem as if it were continuous review to obtain the corresponding (Q, R) policy, and
setting s � R and S � Q � R. We also discussed the application of service levels to
periodic-review systems.
Most real inventory systems involve the management of more than a single product.
We discussed several issues that arise when managing a multiproduct inventory
system. One of these is the amount of time and expense that should be allocated to the
control of each item. The ABC system is a method of classifying items by their annual
volume of sales. Another issue concerns the correct choice of the cost parameters
used in computing inventory control policies. Because many of the cost parameters
used in inventory analysis involve managerial judgment and are not easily measured, it
would be useful if management could compare the effects of various parameter
settings on the performance of the system. A convenient technique for making these
comparisons is via exchange curves. We discussed two of the most popular exchange
curves: (1) the trade-off between the investment in inventory and the frequency of
stock replenishment and (2) the trade-off between the investment in safety stock and
service levels.
nah77856_ch05.indd 296
17/12/14 7:10 PM
nah77856_ch05.qxd
1/17/08
3:53 PM
Page 294
Additional Problems on Stochastic Inventory Models 297
294
Chapter Five Inventory Control Subject to Uncertain Demand
Additional
Problems on
Stochastic
Inventory
Models
28. An artist’s supply shop stocks a variety of different items to satisfy the needs of both
amateur and professional artists. In each case described, what is the appropriate
inventory control model that the store should use to manage the replenishment of the
item described? Choose your answer from the following list and be sure to explain your
answer in each case:
Simple EOQ
Finite production rate
EOQ with quantity discounts
Resource-constrained EOQ
Newsvendor model
Newsvendor model with service level
(Q, R) model with stock-out cost
(Q, R) model with Type 1 service level
(Q, R) model with Type 2 service level
Other type of model
a. A highly volatile paint thinner is ordered once every three months. Cans not
sold during the three-month period are discarded. The demand for the paint
thinner exhibits considerable variation from one three-month period to the
next.
b. A white oil-base paint sells at a fairly regular rate of 600 tubes per month and
requires a six-week order lead time. The paint store buys the paint for $1.20 per
tube.
c. Burnt sienna oil paint does not sell as regularly or as heavily as the white. Sales of
the burnt sienna vary considerably from one month to the next. The useful lifetime
of the paint is about two years, but the store sells almost all the paint prior to the twoyear limit.
d. Synthetic paint brushes are purchased from an East Coast supplier who charges
$1.60 for each brush in orders of under 100 and $1.30 for each brush in orders of
100 or greater. The store sells the brushes at a fairly steady rate of 40 per month for
$2.80 each.
e. Camel hair brushes are purchased from the supplier of part (d), who offers
a discount schedule similar to the one for the synthetic brushes. The camel
hair brushes, however, exhibit considerable sales variation from month to
month.
29. Annual demand for number 2 pencils at the campus store is normally
distributed with mean 1,000 and standard deviation 250. The store purchases the
pencils for 6 cents each and sells them for 20 cents each. There is a two-month
lead time from the initiation to the receipt of an order. The store accountant
estimates that the cost in employee time for performing the necessary paperwork
to initiate and receive an order is $20, and recommends a 22 percent annual interest rate for determining holding cost. The cost of a stock-out is the cost of lost
profit plus an additional 20 cents per pencil, which represents the cost of loss of
goodwill.
a. Find the optimal value of the reorder point R assuming that the lot size used is
the EOQ.
b. Find the simultaneous optimal values of Q and R.
c. Compare the average annual holding, setup, and stock-out costs of the policies
determined in parts (a) and (b).
d. What is the safety stock for this item at the optimal solution?
nah77856_ch05.indd 297
17/12/14 7:10 PM
nah77856_ch05.qxd
1/17/08
3:53 PM
Page 295
298 Chapter Five Inventory Control Subject to Uncertain Demand
Additional Problems on Stochastic Inventory Models 295
30. Consider the problem of satisfying the demand for number 2 pencils faced by the campus store mentioned in Problem 29.
a. Re-solve the problem, substituting a Type 1 service level criterion of 95 percent for
the stock-out cost.
b. Re-solve the problem, substituting a Type 2 service level criterion of 95 percent for
the stock-out cost. Assume that Q is given by the EOQ.
c. Find the simultaneous optimal values of Q and R assuming a Type 2 service level of
95 percent.
31. Answer the following true or false.
a. The lead time is always less than the cycle time.
b. The optimal lot size for a Type 1 service objective of X percent is always less
than the optimal lot size for a Type 2 service objective of X percent for the same
item.
c. The newsvendor model does not include a fixed order cost.
d. ABC analysis ranks items according to the annual value of their demand.
e. For a finite production rate model, the optimal lot size to produce each cycle is equal
to the maximum inventory each cycle.
32. One of the products stocked at Weiss’s paint store, mentioned in Problem 14, is a certain type of highly volatile paint thinner that, due to chemical changes in the product,
has a shelf life of exactly one year. Al Weiss purchases the paint thinner for $20 a gallon
can and sells it for $50 a can. The supplier buys back cans not sold during the year
for $8 for reprocessing. The demand for this thinner generally varies from 20 to 70
cans a year. Al assumes a holding cost for unsold cans at a 30 percent annual interest
rate.
a. Assuming that all values of the demand from 20 to 70 are equally likely, what is the
optimal number of cans of paint thinner for Al to buy each year?
b. More accurate analysis of the demand shows that a normal distribution gives a
better fit of the data. The distribution mean is identical to that used in part (a), and
the standard deviation estimator turns out to be 7. What policy do you now
obtain?
33. Semicon is a start-up company that produces semiconductors for a variety of
applications. The process of burning in the circuits requires large amounts of
nitric acid, which has a shelf life of only three months. Semicon estimates that it
will need between 1,000 and 3,000 gallons of acid for the next three-month
period and assumes that all values in this interval are equally likely. The acid
costs them $150 per gallon. The company assumes a 30 percent annual interest
rate for the money it has invested in inventory, and the acid costs the company $35
a gallon to store. (Assume that all inventory costs are attached to the end of the
three-month period.) Acid that is left over at the end of the three-month period
costs $75 per gallon to dispose of. If the company runs out of acid during the
three-month period, it can purchase emergency supplies quickly at a price of
$600 per gallon.
a. How many gallons of nitric acid should Semicon purchase? Experience with the
marketplace later shows that the demand is closer to a normal distribution, with
mean 1,800 and standard deviation 480.
nah77856_ch05.indd 298
17/12/14 7:10 PM
nah77856_ch05.qxd
296
1/17/08
3:53 PM
Page 296
Chapter Five Inventory Control Subject to Uncertain Demand
Additional Problems on Stochastic Inventory Models 299
b. Suppose that now Semicon switches to a 94 percent fill rate criterion. How
many gallons should now be purchased at the start of each three-month
period?
34. Newsvendor simulator. In order to solve this problem, your spreadsheet program will need
to have a function that produces random numbers [@RAND in Lotus 1–2–3 and
RAND ( ) in Excel]. The purpose of this exercise is to construct a simulation of a periodic-review inventory system with random demand. We assume that the reader is
familiar with the fundamentals of Monte Carlo simulation.
Your spreadsheet should allow for cell locations for storing values of the holding
cost, the penalty cost, the proportional order cost, the order-up-to point, the initial
inventory, and the mean and the standard deviation of periodic demand.
An efficient means of generating an observation from a standard normal variate is
the formula
s
i
Z � [�2 ln(U1)]0.5 cos(2U2),
where U1 and U2 are two independent draws from a (0, 1) uniform distribution. (See,
for example, Fishman, 1973.) Note that two independent calls of @RAND are
required. Since Z is approximately standard normal, the demand X is given by
X � Z � where and are the mean and the standard deviation of one period’s demand.
A suggested layout of the spreadsheet is
NEWSVENDOR SIMULATOR
Holding cost Mean demand Order cost Std. dev. demand Penalty cost Initial inventory Order-up-to point Period
Starting
Inventory
Order
Quantity
Ending
Holding
Demand Inventory
Cost
Penalty
Cost
Order
Cost
1
2
3
.
.
.
20
Totals
Each time you recalculate, the simulator will generate a different sequence of demands. A set of suggested parameters is
h � 2,
c � 5,
p � 20,
� 100,
� 20,
I0 � 50,
Order-up-to point � 150.
nah77856_ch05.indd 299
17/12/14 7:10 PM
nah77856_ch05.qxd
1/17/08
3:53 PM
Page 297
300 Chapter Five Inventory Control Subject to Uncertain Demand
Additional Problems on Stochastic Inventory Models 297
35. Using the simulator for optimization. You will be able to do this problem only if the
program you are using does not restrict the size of the spreadsheet. Assume the
parameters given in Problem 34 but use � 10. Extend the spreadsheet from Problem 34 to 1,000 rows or more and compute the average cost per period. If the
cost changes substantially as you recalculate the spreadsheet, you need to add
additional rows.
Now, experiment with different values of the order-up-to point to find the one that
minimizes the average cost per period. Compare your results to the theoretical optimal
solution.
36. Newsvendor calculator. Design a spreadsheet to calculate the optimal order-up-to point
for a newsvendor problem with normally distributed demands. In order to avoid table
look-ups, use the approximation formula
Z � 5.0633[F 0.135 � (1 � F)0.135]
for the inverse of the standard normal distribution function. (This formula is from
Ramberg and Schmieser, 1972.) The optimal order-up-to point, y*, is of the form
y* � Z � .
a. Assume that all parameters are as given in Problem 34. Graph y* as a function of p,
the penalty cost, for p � 10, 15, . . . , 100. By what percentage does y* increase if p
increases from 10 to 20? from 50 to 100?
b. Repeat part (a) with � 35. Comment on the effect that the variance in demand
has on the sensitivity of y* to p.
37. A large national producer of canned foods plans to purchase 100 combines that are to
be customized for its needs. One of the parts used in the combine is a replaceable blade
for harvesting corn. Spare blades can be purchased at the time the order is placed for
$100 each, but will cost $1,000 each if purchased at a later time because a special production run will be required.
It is estimated that the number of replacement blades required by a combine over
its useful lifetime can be closely approximated by a normal distribution with
mean 18 and standard deviation 5.2. The combine maker agrees to buy back unused
blades for $20 each. How many spare blades should the company purchase with
the combines?
38. Crazy Charlie’s, a discount stereo shop, uses simple exponential smoothing with
a smoothing constant of � .2 to track the mean and the MAD of monthly item demand. One particular item, a stereo receiver, has experienced the following sales over
the last three months: 126, 138, 94.
Three months ago the computer had stored values of mean � 135 and
MAD � 18.5.
a. Using the exponential smoothing equations given in Section 5.1, determine the
current values for the mean and the MAD of monthly demand. (Assume that the stored values were computed prior to observing the demand
of 126.)
b. Suppose that the order lead time for this particular item is 10 weeks
(2.5 months). Assuming a normal distribution for monthly demand, determine
the current estimates for the mean and the standard deviation of lead time
demand.
c. This particular receiver is ordered directly from Japan, and as a result there has
been considerable variation in the replenishment lead time from one order
nah77856_ch05.indd 300
17/12/14 7:10 PM
nah77856_ch05.qxd
298
1/17/08
3:53 PM
Page 298
Chapter Five Inventory Control Subject to Uncertain Demand
39.
40.
41.
42.
Additional Problems on Stochastic Inventory Models 301
cycle to the next. Based on an analysis of past order cycles, it is estimated that the
standard deviation of the lead time is 3.3 weeks. All other relevant figures are
given in parts (a) and (b). Find the mean and the standard deviation of lead time
demand in this case.
d. If Crazy Charlie’s uses a Type 1 service objective of 98 percent to control the
replenishment of this item, what is the value of the reorder level? (Assume that the
lead time demand has a normal distribution.)
The home appliance department of a large department store is using a lot size–reorder
point system to control the replenishment of a particular model of FM table radio. The
store sells an average of 10 radios each week. Weekly demand follows a normal
distribution with variance 26.
The store pays $20 for each radio, which it sells for $75. Fixed costs of replenishment
amount to $28. The accounting department recommends a 20 percent interest rate for the
cost of capital. Storage costs amount to 3 percent and breakage to 2 percent of the value
of each item.
If a customer demands the radio when it is out of stock, the customer will generally
go elsewhere. Loss-of-goodwill costs are estimated to be about $25 per radio. Replenishment lead time is three months.
a. If lot sizes are based on the EOQ formula, what reorder level should be used for the
radios?
b. Find the optimal values of (Q, R).
c. Compare the average annual costs of holding, ordering, and stock-out for the policies that you found in parts (a) and (b).
d. Re-solve the problem using Equations (1) and (2�) rather than (1) and (2). What is
the effect of including lost sales explicitly?
Re-solve the problem faced by the department store mentioned in Problem 39,
replacing the stock-out cost with a 96 percent Type 1 service level.
Re-solve the problem faced by the department store mentioned in Problem 39,
replacing the stock-out cost with a 96 percent Type 2 service level. What is the imputed
shortage cost?
Consider the equation giving the expected average annual cost of the policy (Q, R) in
a continuous-review inventory control system from Section 5.4:
G(Q, R) � ha
Q
K Pn(R)
� r � Tb �
�
.
2
Q
Q
Design a spreadsheet to compute G(Q, R) for a range of values of Q � EOQ
and R � . Use the following approximation formula for L(z) to avoid table
look-ups:
L(z) � exp(�0.92 � 1.19z � 0.37z2).
(This formula is from Parr, 1972.) Store the problem parameters c, h, p, , , and in
cell locations. Visually search through the tabled values of G(Q, R) to discover the
minimum value and estimate the optimal (Q, R) values in this manner. Compare your
results to the true optimal found from manual calculation.
a. Solve Example 5.4.
b. Solve Problem 13 in this manner.
c. Solve Problem 14 in this manner.
nah77856_ch05.indd 301
17/12/14 7:10 PM
nah77856_ch05.qxd
1/17/08
3:53 PM
Page 299
302 Chapter Five Inventory Control Subject to Uncertain Demand
Additional Problems on Stochastic Inventory Models 299
43. The daily demand for a spare engine part is a random variable with a distribution,
based on past experience, given by
Number of
Demands per Day
0
1
2
3
4
Probability
.21
.38
.19
.14
.08
The part is expected to be obsolete after 400 days. Assume that demands from one
day to the next are independent. The parts cost $1,500 each when acquired in
advance of the 400-day period and $5,000 each when purchased on an emergency
basis during the 400-day period. Holding costs for unused parts are based on a daily
interest rate of 0.08 percent. Unused parts can be scrapped for 10 percent of their
purchase price. How many parts should be acquired in advance of the 400-day
period? (Hint: Let D1, D2, . . . , D400 represent the daily demand for the part. Assume
each Di has mean and variance 2. The central limit theorem says that the total
demand for the 400-day period, Di , is approximately normally distributed with
mean 400 and variance 4002.)
44. Cassorla’s Clothes sells a large number of white dress shirts. The shirts, which bear
the store label, are shipped from a manufacturer in New York City. Hy Cassorla, the
proprietor, says, “I want to be sure that I never run out of dress shirts. I always try to
keep at least a two months’ supply in stock. When my inventory drops below that
level, I order another two-month supply. I’ve been using that method for 20 years, and
it works.”
The shirts cost $6 each and sell for $15 each. The cost of processing an order and
receiving new goods amounts to $80, and it takes three weeks to receive a shipment.
Monthly demand is approximately normally distributed with mean 120 and standard
deviation 32. Assume a 20 percent annual interest rate for computing the holding
cost.
a. What value of Q and R is Hy Cassorla using to control the inventory of white dress
shirts?
b. What fill rate (Type 2 service level) is being achieved with the current policy?
c. Based on a 99 percent fill rate criterion, determine the optimal values of Q
and R that he should be using. (Assume four weeks in a month for your
calculations.)
d. Determine the difference in the average annual holding and setup costs between the
policies in parts (b) and (c).
e. Estimate how much time would be required to pay for a $25,000 inventory control
system, assuming that the dress shirts represent 5 percent of Hy’s annual business and
that similar savings could be realized on the other items as well.
45. The Poisson distribution is discussed in Appendix 5–D at the end of this chapter.
Assume that the distribution of bagels sold daily at Billy’s Bakery in Problem 8
follows a Poisson distribution with mean 16 per day. Using Table A–3 in the back of
the book or the Poisson distribution function built into Excel, determine the optimal
number of bagels for Billy’s to bake each day.
nah77856_ch05.indd 302
17/12/14 7:10 PM
nah77856_ch05.qxd
1/17/08
3:53 PM
Page 300
300 Chapter Five Inventory Control Subject to Uncertain Demand
Additional Problems on Stochastic Inventory Models 303
46. Consider the Crestview Printing Company described in Problem 9. Suppose that
sales of cards (in units of 50,000) follow a Poisson distribution with mean 6.3. Using
Table A–3 in the back of the book or the Poisson distribution function built into
Excel find the optimal number of cards for Crestview to print for the next Christmas
season.
47. The Laplace distribution is discussed in Appendix 5–D. As noted there, the Laplace
distribution could be a good choice for describing demand for slow-moving items
and for fast-moving items with high variation. The cdf of the Laplace distribution is
given by
F(x) � 0.5[1 � sgn (x � )](1 � exp (�|x � |>),
and the inverse of the cdf is given by
F �1( p) � � sgn ( p � 0.5) ln (1 � 2|p � 0.5|),
48.
49.
50.
51.
where sgn(x) is the sign of x.The mean is m and the variance is 22. Since the inverse distribution function can be written down explicitly, one does not have to resort to tables to
solve newsvendor problems when demand follows the Laplace distribution.
Solve Problem 10, part (a), assuming the demand for the EX123 follows a Laplace
distribution with parameters � 60 and � 322 (which will give exactly the same
mean and variance).
Solve Problem 11 assuming the demand for fans over the selling season follows a
Laplace distribution with the same mean and variance as you computed in Problem
11(a).
Solve Problem 12(b) assuming the demand for handbags over the selling season
follows a Laplace distribution with mean 150 and standard deviation 20.
Solve Problem 13 assuming that the lead time demand follows a Laplace distribution
with mean and variance equal to that which was computed in Problem 13. What difference do you see in the (Q, R) values as compared to those for the normal case?
Solve Problem 14 assuming that the lead time demand follows a Laplace distribution
with mean and variance equal to the mean and variance of lead time demand you
computed for Problem 14. What difference do you see in the (Q, R) values as
compared to those for the normal case?
Appendix 5–A
Notational Conventions and Probability Review
Demand will be denoted by D, which is assumed to be a random variable. The cumulative
distribution function (cdf) of demand is F(x) and is defined by
F(x) � P{D … x}
for �q 6 x 6 �q.
When D is continuous, the probability density function (pdf) of demand, f(x), is
defined by
dF(x)
.
f (x) �
dx
When D is discrete, f(x) is the probability function (pf) defined by
f(x) � PX � x � F(x) � F(x � 1).
nah77856_ch05.indd 303
17/12/14 7:10 PM
nah77856_ch05.qxd
1/17/08
3:53 PM
Page 301
Appendix 5–B 301
304 Chapter Five Inventory Control Subject to Uncertain Demand
Note that in the continuous case the density function is not a probability and the value of
f(x) is not necessarily less than 1, although it is always true that f(x) � 0 for all x.
The expected value of demand, E(D), is defined as
�q
E(D) �
in the continuous case, and
L�q
xf (x) dx
E(D) � a xf(x)
q
�q
x��
in the discrete case.
We use the symbol to represent the expected value of demand [E(D) � ]. In what
follows we assume that D is continuous; similar formulas hold in the discrete case. Let g(x)
be any real-valued function of the real variable x. Then
�q
E(g(D)) �
L�q
g(x) f (x)dx.
In particular, let g(D) � max(0, Q � D). Then
�q
E(g(D)) �
L�q
max(0, Q � x) f (x) dx.
Because demand is nonnegative, it must be true that f(x) � 0 for x � 0. Furthermore, when
x � Q, max(0, Q � x) � 0, so we may write
E(g(D)) �
L0
Q
(Q � x) f (x)dx.
In the analysis of the newsvendor model, Leibniz’s rule is used to determine the derivative
of G(Q). According to Leibniz’s rule:
a (y)
a (y)
2
2
d
h(x, y) dx �
[�h(x, y)�y] dx � h(a2(y), y)a�2(y) � h(a1(y), y)a�1(y).
dy La1(y)
La1(y)
Appendix 5–B
Additional Results and Extensions
for the Newsvendor Model
1. INTERPRETATION OF THE OVERAGE AND UNDERAGE
COSTS FOR THE SINGLE PERIOD PROBLEM
Define
S � Selling price of the item.
c � Variable cost of the item.
h � Holding cost per unit of inventory remaining in stock at the end of the
period.
p � Loss-of-goodwill cost plus bookkeeping expense (charged against the number of
back orders on the books at the end of the period).
nah77856_ch05.indd 304
17/12/14 7:10 PM
nah77856_ch05.qxd
1/17/08
3:53 PM
Page 302
302 Chapter Five Inventory Control Subject to Uncertain Demand
Appendix 5–B 305
We show how cu and co should be interpreted in terms of these parameters. As earlier, let Q
be the order quantity and D the demand during the period. Assume without loss of generality
that starting inventory is zero. Then the cost incurred at the end of the period is
cQ � h max(Q � D, 0) � p max(D � Q, 0) � S min(Q, D).
The expected cost is
G(Q) � cQ � h
�S
Using
L0
Q
xf(x) dx �
L0
L0
L0
Q
(Q � x)f(x) dx � p
Q
xf(x) dx � SQ
q
xf(x) dx �
LQ
LQ
LQ
q
(x � Q)f(x) dx
q
f(x) dx.
q
xf(x) dx � �
LQ
q
xf(x) dx,
the expected cost may be written
G(Q) � cQ � h
L0
Q
(Q � x)f(x) dx � (p � S)
LQ
q
(x � Q)f(x) dx � S.
The optimal order quantity satisfies
G′(Q) � 0
or
c � hF(Q) � (p � S)(1 � F(Q)) � 0,
which results in
F(Q) �
P�S�c
.
p�S�h
Setting cu � p � S � c and co � h � c gives the critical ratio in the form
cu(cu � co).
2. The Newsvendor Cost When Demand is Normal
2.
EXTENSION
TO INFINITE
HORIZON
ASSUMING
When
the one period demand
follows a normal
distribution,
we can obtain an explicit
­expression
for the optimal one period
costDEMAND
for the newsvendor model from Section 5.3.
FULL
BACK-ORDERING
OF
We know that the expected single period cost function is
Let D1, D2, . . . be an infinite sequence
of demands. Assume
that the demands are
Q

independent identicallyG(Q)
distributed
random
variables
having
common
(t - distribution
Q) f (t) dt, function
(Q - t) f(t) dt + cu
= co
F(x) and density f(x). The policy isL
to
LQ As all excess demand is
0 order up to Q each period.
and the optimal
Q satisfies
back-ordered,
thevalue
order of
quantity
in any period is exactly the previous period’s demand. The
number of units sold will also equal the demand.
In
cuorder to see this, consider the number
*
of units sold in successive periods: F( Q ) = c + c .
u
o
Number of units sold in period 1 � min(Q, D1),
It is convenient to express G(Q) in a slightly different form. Note that
Number
of units sold in period 2 � max(D1 � Q, 0) � min(Q, D2),
Q
Number
units
period
� dt
max(D
min(Q,
(Q - t)off (t)
dt sold
= in(Q
- t) 3f (t)
- 2�
(QQ,- 0)t) �
f (t)
dt = QD3-), m + n(Q) where
L0
L0
LQ
and so on.

(t - Q) f (t) dt.
n(Q) =
LQ
nah77856_ch05.indd 305
17/12/14 7:10 PM
306 Chapter Five Inventory Control Subject to Uncertain Demand
It follows that G(Q) can be written in the form
G(Q) = co (Q - m) + (cu + co)n(Q).
where μ is the expected demand.
We know from Section 5.4 that one can write an explicit expression for n(Q) when demand
is normal. In particular, we showed that n(Q) = sL(z) where L(z) is the standard loss
integral. As noted in Section 5.4,
L(z) = f(z) - z[1 - f(z)].
where f(z) is the standard normal density, f(z) the standard normal distribution
­function and z = (Q - m)>s is the standardized value of the order quantity Q.
It follows that in the normal case we can write
G(Q) = co (Q - m) + (cu + co)s[f(z) - z(1 - f(z)].
The standard normal density function can be computed directly, and the cumulative
­distribution function is available through tables and is a built in function in Excel.
It might also be of interest to know the minimum value of the expected cost. This
is the value of the function G when Q = Q*. Note that at Q = Q* the cumulative
distribution function f(z*) is equal to the critical ratio cu >(cu + co), which means
the complementary cumulative distribution function 1 - f(z*) is equal to the ratio
co >(cu + co). Also, since z = (Q - m)>s, it follows that Q - m = sz. Making
these two substitutions in the expression above for G(Q) we obtain
co
G(Q*) = cosz* + (cu + co)s c f(z*) - z* a
b d = (cu + co)sf(z*).1
co + cu
Note that if one knows z* no table look up is required to compute this expression since
we know that
1
2
f(z) =
e - 0.5z .
22p
The expression for G(Q*) shows that the optimal newsvendor cost increases linearly
in both the underage and overage costs as well as in the standard deviation of demand.
3. Extension to Infinite Horizon Assuming
Full Back-Ordering of Demand
Let D1, D2, . . . be an infinite sequence of demands. Assume that the demands are
­independent identically distributed random variables having common distribution
function F(x) and density f(x). The policy is to order up to Q each period. As all excess demand is back-ordered, the order quantity in any period is exactly the previous
­period’s demand. The number of units sold will also equal the demand. In order to see
this, consider the number of units sold in successive periods:
Number of units sold in period 1 5 min(Q, D1),
Number of units sold in period 2 5 max(D1 2 Q, 0) 1 min(Q, D2),
Number of units sold in period 3 5 max(D2 2 Q, 0) 1 min(Q, D3),
and so on.
1
Porteus (2002) has derived essentially the same expression on page 13. I am grateful to Gerard Cachon
for helpful discussions on this point.
nah77856_ch05.indd 306
17/12/14 7:10 PM
Appendix 5–B 307
These relationships follow because back-ordering of excess demand means that the
sales are made in the subsequent period. Now, notice that
min(Q, Di) 1 max(Di 2 Q, 0) 5 Di for i 5 1, 2, . . . ,
which follows from considering the cases Q , Di and Q $ Di.
Hence, the expected cost over n periods is
cQ + (c - S)E(D1 + D2 + c + Dn - 1) - (S)E[min(Q, Dn)] + nL (Q)
where
= cQ + (c - S)(n - 1)m - (S)E[min(Q, Dn)] + nL(Q)
L(Q) = h
L0
Q
(Q - x)f(x)dx + p
LQ

(x - Q)f(x) dx.
Dividing by n and letting n → q gives the average cost over infinitely many
­periods as
(c - S)m + L(Q).
The optimal value of Q occurs where L9(Q) 5 0, which results in
F(Q) = p / (p + h).
4. Extension to Infinite Horizon Assuming
Lost Sales
If excess demand is lost rather than back-ordered, then the previous argument is no
longer valid. The number of units sold in period 1 is min(Q, D1 ), which is also the
number of units ordered in period 2; the number of units sold in period 2 is min(Q, D2)
(since there is no back-ordering of excess demand), which is also the number of units
ordered in period 3; and so on. As shown in Section 1 of this appendix,

(x - Q)f(x) dx.
LQ
Hence, it follows that the expected cost over n periods is given by
E[min(Q, D)] = m -
If we divide by n and let n →
cost per period:

(x - Q)f(x) dx d + nL(Q)
LQ
, we obtain the following expression for the average
cQ + [(n - 1)c - nS] c m 
(x - Q)f(x) dx d + L(Q).
LQ
Differentiating with respect to Q and setting the result equal to zero gives the
­following condition for the optimal Q:
(c - S)c m -
F(Q) =
p + S - c
.
p + S + h - c
Setting cu = p + S - c and c = h gives the critical ratio in the form cu / (cu + c).

Hence, we interpret cu as the cost of the loss of goodwill plus the lost profit per sale, and
co as the cost of holding only.
nah77856_ch05.indd 307
17/12/14 7:10 PM
nah77856_ch05.qxd
1/17/08
3:53 PM
Page 304
304 Chapter Five Inventory
Five Inventory Control
Control Subject
Subject to
to Uncertain
Uncertain Demand
Demand
308 Chapter
Appendix 5–C
Derivation of the Optimal (Q, R) Policy
From Section 5.4, the objective is to find values of the variables Q and R to minimize the
function
G(Q, R) � h(Q2 � R � ) � KQ � pn(R)Q.
(1)
Because this function is to be minimized with respect to the two variables (Q, R), a necessary condition for optimality is that �G�Q � �G�R � 0. The two resulting equations are
0G h K pn(R)
� �
�
� 0,
0Q 2 Q2
Q2
0G
� h � pn¿(R)>Q � 0.
0R
Note that since n(R) �
LR
(2)
(3)
q
(x � R) f (x) dx, one can show that
n(R) � �(1 � F(R)).
From Equation (2) we obtain
or
1
h
[K � pn(R)] �
2
2
Q
Q2 �
2K � 2pn(R)
,
h
which gives
Q�
From Equation (3) we obtain
2[K � pn(R)]
.
A
h
(4)
h � p[�(1 � F(R))]Q � 0,
which gives
1 � F(R) � Qhp.
(5)
Author’s note: We use the term optimal somewhat loosely here. Technically speaking, the
model that gives rise to these two equations is only approximate for several reasons. For one,
the use of the average expected inventory is an approximation, since this is not the same as
the expected average inventory (because one should not charge holding costs when inventory
goes negative). As we saw in the discussion of negative safety stock in this chapter, there are
cases where the right-hand side of Equation (5) can exceed 1, and the model “blows up.”
This would not be the case for a truly exact model, which is beyond the scope of this book.
Appendix 5–D
Probability Distributions for Inventory Management
In this chapter we have made frequent reference to the normal distribution as a model for
demand uncertainty. Although the normal distribution certainly dominates applications, it is
not the only choice available. In fact, it could be a poor choice in some circumstances. In this
appendix we discuss other distributions for modeling demand uncertainty.
nah77856_ch05.indd 308
17/12/14 7:10 PM
nah77856_ch05.qxd
1/17/08
3:53 PM
Page 305
Appendix
305
Appendix 5–D
5–D 309
1. THE POISSON DISTRIBUTION AS A MODEL
OF DEMAND UNCERTAINTY
One situation in which the normal distribution may be inappropriate is for slow-moving
items, that is, ones with small demand rates. Because the normal is an infinite distribution,
when the mean is small it is possible that a substantial portion of the density curve extends
into the negative half line. This could give poor results for safety stock calculations. A common choice for modeling slow movers is the Poisson distribution. The Poisson is a discrete
distribution defined on the positive half line only. Let X have the Poisson distribution with
parameter . Then
e�x
f(x) �
for x � 0, 1, 2, . . . .
x!
(The derivation of the Poisson distribution and its relationship to the exponential
distribution are discussed in detail in Section 12.3).
An important feature of the Poisson distribution is that both the mean and the variance are
equal to (giving � ). Hence, it should be true that the observed standard deviation
of periodic demand (or lead time demand) is close to the square root of the mean periodic
demand (or lead time demand) for the Poisson to be an appropriate choice. Table A–3 at the
back of the book is a table of the complementary cumulative Poisson distribution. This table
allows one to compute optimal policies for the newsvendor and the (Q, R) models assuming that
demand follows the Poisson distribution.
We show how to find optimal policies for both the newsvendor and the (Q, R) models when
demand is Poisson. For the newsvendor case, one simply applies the method outlined in Section 5.3 for discrete demand and obtains the probabilities from Table A–3. We illustrate
with an example.
Example 5D.1
Consider Mac’s newsstand, discussed in several examples in Chapter 5. A few regular customers
have asked Mac to stock a particular stereo magazine. Mac has agreed even though sales have
been slow. Based on past data, Mac has found that a Poisson distribution with mean 5 closely
fits the weekly sales pattern for the magazine. Mac purchases the magazines for $0.50 and sells
them for $1.50. He returns unsold copies to his supplier, who pays Mac $0.10 for each. Find the
number of these magazines he should purchase from his supplier each week.
Solution
The overage cost is co � $0.50 � $0.10 � $0.40 and the underage cost is cu � $1.50 � $0.50 �
$1.00. It follows that the critical ratio is cu(co � cu) � 11.4 � .7143. As Table A–3 gives the
complementary cumulative distribution, we subtract table entries from 1 to obtain values of the
cumulative distribution function. From the table we see that
F(5) � 1 � .5595 � .4405,
F(6) � 1 � .3840 � .6160,
F(7) � 1 � .2378 � .7622.
Because the critical ratio is between F(6) and F(7), we move to the larger value, thus giving
an optimal order quantity of 7 magazines.
It is not difficult to calculate optimal (Q, R) policies when the demand follows a Poisson distribution, but the process requires an iterative solution similar to that for the normal distribution described in Section 5.5. Define P(x) as the complementary cumulative probability of the
Poisson. That is,
q
P(x) � a f(k).
k� x
Table A–3 gives values of P(x). It can then be shown that for the Poisson distribution
n(R) � mP(R) � RP(R � 1),
nah77856_ch05.indd 309
17/12/14 7:10 PM
nah77856_ch05.qxd
1/17/08
3:53 PM
Page 306
306
Chapter Five
Inventory Control Subject to Uncertain Demand
310 Chapter
Five Inventory
which means that all the information required to compute optimal policies appears in Table A–3.1
That is, special tables for computing n(R) are not required in this case. This relationship allows one
to compute the (Q, R) policy using the pair of equations (1) and (2) from Section 5.4.
Example 5D.2
A department store uses a (Q, R) policy to control its inventories. Sales of a pocket FM radio
average 1.4 radios per week. The radio costs the store $50.00, and fixed costs of replenishment
amount to $20.00. Holding costs are based on a 20 percent annual interest charge, and the
store manager estimates a $12.50 cost of lost profit and loss of goodwill if the radio is demanded when out of stock. Reorder lead time is 2 weeks, and statistical studies have shown
that demand over the lead time is accurately described by a Poisson distribution. Find the optimal
lot sizes and reorder points for this item.
Solution
The relevant parameters for this problem are
� (1.4)(52) � 72.8.
h � (50)(0.2) � 10.0.
K � $20.
� (1.4)(2) � 2.8.
p � $12.50.
To start the solution process, we compute the EOQ. It is
EOQ �
2K
B h
The next step is to find R solving
�
B
(2)(20)(72.8)
� 17.1.
10
P(R) � Qhp.
Substituting the EOQ for Q and solving gives P(R0) � .1868. From Table A–3, we see that
R � 4 results in P(R) � .3081 and R � 5 results in P(R) � .1523. Assuming the conservative strategy of rounding to the larger R, we would choose R � 5 and P(R) � .1523. It now follows that
n(R0) � (2.8)(.1523) � (5)(.0651) � 0.1009.
Hence Q1 is
Q1 �
B
(2)(72.8)[20 � (12.5)(0.1009)]
� 17.6.
10
It follows that P(R1) � .1934, giving R1 � R0 � 5. Hence the solution has converged, because
successive R (and, hence, Q) values are the same. The optimal solution is (Q, R) � (18, 5).
2. THE LAPLACE DISTRIBUTION
A continuous distribution, which has been suggested for modeling slow-moving items or
ones with more variance in the tails than the normal, is the Laplace distribution. The Laplace
distribution has been called the pseudo-exponential distribution, as it is mathematically an
exponential distribution with a symmetric mirror image. (The exponential distribution is
discussed at length in Chapter 12 in the context of reliability management.)
The mathematical form of the Laplace pdf is
f(x) �
1
exp(��x � �)
2
for �q 6 x 6 �q.
Because the pdf is symmetric around , the mean is . The variance is 22. The
Laplace distribution is also a reasonable model for slow-moving items, and is an
1
nah77856_ch05.indd 310
See Hadley and Whitin (1963), p. 441.
17/12/14 7:10 PM
nah77856_ch05.qxd
1/17/08
3:53 PM
Page 307
Appendix 5–D
307
5–D 311
alternative to the normal distribution for fast-moving items when there is more spread in
the tails of the distribution than the normal distribution gives. As far as inventory
applications are concerned, Presutti and Trepp (1970) noticed that it significantly simplified
the calculation of the optimal policy for the (Q, R) model.
One can show that for any value of R � �, the complementary cumulative distribution
(P(R)) and the loss integral (n(R)) are given by
P(R) � 0.5 exp(�[(R � �)�])
n(R) � 0.5� exp(�[(R � �)�])
so that the ratio n(R)P(R) � �, independent of R. This fact results in a very simple solution
for the (Q, R) model. Recall that the two equations defining the optimal policy were
Q�
A
2[K � pm(R)]
h
P(R) � Qhp�
[where P(R) � 1 � F(R)].
The simplification is achieved by using the SOQ formula presented in Section 5.5. The
SOQ representation is an alternative representation for Q that does not include the stockout cost, p. Using P(R) � 1 � F(R), the SOQ formula is
Q�
n(R)
n(R) 2
2K�
�
�a
b
P(R) A h
P(R)
2K
� �2
A h
independently of R! Hence, the optimal Q and R can be found in a simple one-step
calculation. When using a cost model, find the value of P(R) from P(R) � Qhp�. Then,
using the representation P(R) � exp[�(R � �)�)]2, it follows that R � �� ln (2P(R))
� �. Using a service level model, one simply uses the formulas for R given in Section 5.5.
We illustrate with an example.
���
Example 5D.3
Consider Example 5D.2, but suppose that we wish to use the Laplace distribution to compute
the optimal policy rather than the Poisson distribution. As both the mean and the variance of
the lead time demand are 2.8, we set � � 2.8 and 2�2 � 2.8, giving � � 1.1832. It follows
that
Q � 1.1832 �
(2)(20)(72.8)
� (1.1832)2 � 18.3.
10
B
As with Example 5D.2, we obtain P(R) � .1934. Using R � �� ln(2P(R)) � � and substituting P(R) � .1934 gives R � 3.92, which we round to 4. Notice that this solution differs slightly
from the one we obtained assuming Poisson demand. However, recall that using the Poisson
distribution we found that the optimal R was between 4 and 5, which we chose to round to 5
to be conservative.
3. OTHER LEAD TIME DEMAND DISTRIBUTIONS
Many other probability distributions have been recommended for modeling lead time
demand. Some of these include the negative binomial distribution, the gamma distribution, the logarithmic distribution, and the Pearson distribution. (See Silver and Peterson,
nah77856_ch05.indd 311
17/12/14 7:10 PM
nah77856_ch05.qxd
1/17/08
3:53 PM
Page 308
308
Chapter Five
Inventory Control
312 Chapter
Five Inventory
Control Subject
Subject to
to Uncertain
Uncertain Demand
Demand
1985, p. 289, for a list of articles that discuss these distributions in the context of inventory management.) The normal distribution probably accounts for the lion’s share of
applications, with the Poisson accounting for almost all the rest. We included the Laplace
distribution because of the interesting property that optimal (Q, R) policies can be found
without an iterative solution, and because it could be a good alternative to the Poisson for
low-demand items.
Appendix 5–E
Glossary of Notation for Chapter 5
Desired level of Type 1 service.
Desired level of Type 2 service.
co Unit overage cost for newsvendor model.
cu Unit underage cost for newsvendor model.
D Random variable corresponding to demand. It is the one-period
demand for the newsvendor model and the lead time demand for the
(Q, R) model.
EOQ Economic order quantity.
F(t) Cumulative distribution function of demand. Values of the standard
normal CDF appear in Table A–4 at the end of the book.
f(t) Probability density function of demand.
G(Q) Expected one-period cost associated with lot size Q (newsvendor model).
G(Q, R) Expected average annual cost for the (Q, R) model.
h Holding cost per unit per unit time.
I Annual interest rate used to compute holding cost.
K Setup cost or fixed order cost.
Expected demand rate (units per unit time).
L(z) Normalized loss function. Used to compute n(R) L(z). Tabled
values of L(z) appear in Table A–4 at the end of the book.
Mean demand [lead time demand for (Q, R) model].
n(R) Expected number of stock-outs in the lead time for (Q, R) model.
p Penalty cost per unit for not satisfying demand.
Q Lot size or size of the order.
S Safety stock; S R for (Q, R) model.
SOQ Service order quantity.
T Expected cycle time; mean time between placement of successive
orders.
Order lead time.
Type 1 service Proportion of cycles in which all demand is satisfied.
Type 2 service Proportion of demands satisfied.
nah77856_ch05.indd 312
17/12/14 7:11 PM
nah77856_ch05.qxd
1/9/08
11:42 AM
Page 309
Bibliography 309
Bibliography 313
Bibliography
Agrawal, N., and S. A. Smith, Retail Supply Chain
Management. Springer, 2008.
Arrow, K. A.; T. E. Harris; and T. Marschak. “Optimal
Inventory Policy.” Econometrica 19 (1951), pp. 250–72.
Arrow, K. A.; S. Karlin; and H. E. Scarf, eds. Studies in
the Mathematical Theory of Inventory and Production.
Stanford, CA: Stanford University Press, 1958.
Bessler, S. A., and A. F. Veinott, Jr. “Optimal Policy for a
Dynamic Multiechelon Inventory Model.” Naval Research
Logistics Quarterly 13 (1966), pp. 355–89.
Brown, R. G. Statistical Forecasting for Inventory Control.
New York: McGraw-Hill, 1959.
Brown, R. G. Decision Rules for Inventory Management.
Hinsdale, IL: Dryden Press, 1967.
Clark, A., and H. E. Scarf. “Optimal Policies for a
Multiechelon Inventory Problem.” Management Science 6
(1960), pp. 475–90.
Cohen, M. A., and D. Pekelman. “LIFO Inventory Systems.”
Management Science 24 (1978), pp. 1150–62.
Deuermeyer, B. L., and L. B. Schwarz. “A Model for the
Analysis of System Service Level in Warehouse-Retailer
Distribution Systems: Identical Retailer Case.” In Multilevel
Production/Inventory Control Systems: Theory and Practice,
ed. L. B. Schwarz, pp. 163–94. Amsterdam: North Holland,
1981.
Dvoretsky, A.; J. Kiefer; and J. Wolfowitz. “The Inventory
Problem: I. Case of Known Distributions of Demand.”
Econometrica 20 (1952), pp. 187–222.
Ehrhardt, R. “The Power Approximation for Computing
(s, S) Inventory Policies.” Management Science 25 (1979),
pp. 777–86.
Eppen, G., and L. Schrage. “Centralized Ordering Policies
in a Multi-Warehouse System with Lead Times and Random
Demand.” In Multilevel Production/Inventory Control
Systems: Theory and Practice, ed. L. B. Schwarz, pp. 51–68.
Amsterdam: North Holland, 1981.
Federgruen, A., and P. Zipkin. “Computational Issues in
an Infinite Horizon Multi-Echelon Inventory Model.”
Operations Research 32 (1984), pp. 818–36.
Fishman, G. S. Concepts and Methods in Discrete Event
Digital Simulation. New York: John Wiley & Sons, 1973.
Freeland, J. R., and E. L. Porteus. “Evaluating the
Effectiveness of a New Method of Computing
Approximately Optimal (s, S) Inventory Policies.”
Operations Research 28 (1980), pp. 353–64.
Graves, S. C. “A Multiechelon Inventory Model for a
Repairable Item with One for One Replenishment.”
Management Science 31 (1985), pp. 1247–56.
Graves, S. C.; A. H. G. Rinnooy Kan; and P. Zipkin, eds.
Handbooks in Operations Research and Management
nah77856_ch05.indd 313
Science. Volume 4, Logistics of Production and Inventory.
Amsterdam: Elsevier Science Publishers, 1992.
Hadley, G. J., and T. M. Whitin. Analysis of Inventory
Systems. Englewood Cliffs, NJ: Prentice Hall, 1963.
Hartung, P. “A Simple Style Goods Inventory Model.”
Management Science 19 (1973), pp. 1452–58.
Hausman, W. H., and R. Peterson. “Multiproduct Production
Scheduling for Style Goods with Limited Capacity, Forecast
Revisions, and Terminal Delivery.” Management Science 18
(1972), pp. 370–83.
Iglehart, D. L., “Optimality of (s, S) Inventory Policies in the
Infinite Horizon Dynamic Inventory Problem.” Management
Science 9 (1963), pp. 259–67.
Kaplan, R. “A Dynamic Inventory Model with Stochastic
Lead Times.” Management Science 16 (1970), pp. 491–507.
Love, S. F. Inventory Control. New York: McGraw-Hill, 1979.
Mohan, S. “EDI’s Move to Prime Time Stalled by Cost
Perception.” Computerworld 29 (February 20, 1995), p. 91.
Muckstadt, J. M., and L. J. Thomas. “Are Multiechelon
Inventory Models Worth Implementing in Systems with
Low Demand Rate Items?” Management Science 26 (1980),
pp. 483–94.
Murray, G. R., and E. A. Silver. “A Bayesian Analysis of the
Style Goods Inventory Problem.” Management Science 12
(1966), pp. 785–97.
Murray, J. E. “The EDI Explosion.” Purchasing 118
(February 16, 1995), pp. 28–30.
Nahmias, S. “Inventory Models.” In The Encyclopedia of
Computer Science and Technology, Volume 9, ed. J. Belzer,
A. G. Holzman, and A. Kent, pp. 447–83. New York: Marcel
Dekker, 1978.
Nahmias, S. “Simple Approximations for a Variety of
Dynamic Lead Time Lost-Sales Inventory Models.”
Operations Research 27 (1979), pp. 904–24.
Nahmias, S. “Managing Reparable Item Inventory Systems:
A Review.” In Multilevel Production/Inventory Control Systems: Theory and Practice, ed. L. B. Schwarz, pp. 253–77.
Amsterdam: North Holland, 1981.
Nahmias, S. “Perishable Inventory Theory: A Review.”
Operations Research 30 (1982), pp. 680–708.
Nahmias, S., and S. Smith. “Mathematical Models of
Retailer Inventory Systems: A Review.” In Perspectives in
Operations Management: Essays in Honor of Elwood
S. Buffa, ed. R. K. Sarin. Boston: Kluwer, 1992.
Parr, J. O. “Formula Approximations to Brown’s Service
Function.” Production and Inventory Management 13
(1972), pp. 84–86.
Porteus, E.
E. L.
L. “Numerical
FoundationsComparisons
of StochasticofInventory
Porteus,
InventoryTheory.
Policies
Stanford,
CA:
Stanford
Business
Books, 2002.
for
Periodic
Review
Systems.”
Operations
Research 33
(1985), pp. 134–52.
17/12/14 7:11 PM
36.
Discrete Event
y & Sons, 1973.
uating the
puting
Policies.”
–64.
Model for a
nishment.”
–56.
d P. Zipkin, eds.
Management
Nahmias, S. “Perishable Inventory Theory: A Review.”
Operations Research 30 (1982), pp. 680–708.
Nahmias, S., and S. Smith. “Mathematical Models of
Retailer Inventory Systems: A Review.” In Perspectives in
Operations Management: Essays in Honor of Elwood
S. Buffa, ed. R. K. Sarin. Boston: Kluwer, 1992.
Parr, J. O. “Formula Approximations to Brown’s Service
314 Chapter
Five Inventory
Control Subject
to Uncertain
Function.”
Production
and Inventory
Management
13 Demand
(1972), pp. 84–86.
Porteus, E. L. “Numerical Comparisons of Inventory Policies
for Periodic Review Systems.” Operations Research 33
(1985), pp. 134–52.
Presutti, V., and R. Trepp. “More Ado about EOQ.” Naval
Research Logistics Quarterly 17 (1970), pp. 243–51.
Ramberg, J. S., and B. W. Schmeiser. “An Approximate
Method for Generating S
­ ymmetric Random Variables.” Communications of the ACM 15 (1972), pp. 987–89.
Scarf, H. E. “The Optimality of (s, S) Policies in the
­Dynamic Inventory Problem.” In Mathematical Methods
in the Social Sciences, ed. K. J. Arrow, S. Karlin, and
P. Suppes. Stanford, CA: Stanford University Press, 1960.
Scarf, H. E. “Analytical Techniques in Inventory Theory.”
In Multi-Stage Inventory Models and Techniques, ed.
H. E. Scarf, D. M. Gilford, and M. W. Shelly. S
­ tanford, CA:
­Stanford University Press, 1963.
nah77856_ch05.indd 314
Schmidt, C. P., and S. Nahmias. “Optimal Policy for a
­Single Stage Assembly ­System with Stochastic Demand.”
Operations Research 33 (1985), pp. 1130–45.
Sherbrooke, C. C. “METRIC: Multiechelon Technique for
Recoverable Item C
­ ontrol.” Operations Research 16 (1968),
pp. 122–41.
Silver, E. A., and R. Peterson. Decision Systems for Inventory
Management and ­Production Planning. 2nd ed. New York:
John Wiley & Sons, 1985.
Veinott, A. F. “The Status of Mathematical Inventory
­Theory.” Management Science 12 (1966), pp. 745–77.
Veinott, A. F., and H. M. Wagner. “Computing Optimal
(s, S) Inventory Policies.” Management Science 11 (1965),
pp. 525–52.
Whitin, T. M. The Theory of Inventory Management.
Rev. ed. Princeton, NJ: P
­ rinceton University Press, 1957.
17/12/14 7:11 PM
Chapter Six
Supply Chain
­Management
“Supply chains cannot tolerate even 24 hours of disruption. So if you lose your place in
the supply chain because of wild behavior, you could lose a lot. It would be like pouring
cement down one of your oil wells.”
—Thomas Friedman
Chapter Overview
Purpose
To understand what a modern supply chain is, how supply chains are organized and
managed, and to review the newest developments in this important area.
Key Points
1. What is a supply chain? A supply chain is the entire network comprising the activities of a firm that links suppliers, factories, warehouses, stores, and customers.
It requires management of goods, money, and information among all the relevant
players. While the specific term supply chain management (SCM) emerged only
in the late 1980s, managing the flow of goods has been an issue since the industrial revolution, and was traditionally simply called logistics.
2. Supply chain strategy. For most products it is not possible to have a supply chain
that is both low cost and highly responsive. A supply chain strategy must therefore align with the product’s positioning in the marketplace. In particular, a product that competes on price must be delivered through a highly efficient supply
chain, while for an innovative or high-fashion item it is most important to be able
to respond quickly to changes in customer demand.
3. The role of information in supply chains. As has been noted, a supply chain
involves the transfer of goods, money, and information. Modern supply chain
management seeks to eliminate the inefficiencies that arise from poor information flows. One way to ameliorate this problem is vendor-managed inventories,
where vendors, rather than retailers, are responsible for keeping inventory on the
shelves. Advances in technology have also improved the availability of information in supply chains.
4. The transportation problem. The transportation problem is one of the early
­applications of linear programming. Assume m production facilities (sources) and n
demand points (sinks). The unit cost of shipping from every source to every sink is
known, and the objective is to determine a shipping plan that satisfies the ­supply
315
nah77856_ch06.indd 315
17/12/14 6:55 PM
316 Chapter Six Supply Chain Management
and demand constraints at minimum cost. The linear programming ­formulation
of the transportation problem has been successfully solved with hundreds of
­thousands of variables and constraints. A generalization of the transportation
problem is the transshipment problem, where intermediate nodes can be used for
storage as well as be demand or supply points. Transshipment problems can also
be solved with linear programming.
5. Routing in supply chains. Consider a delivery truck that must make deliveries to
several customers. The objective is to find the optimal sequence of deliveries that
minimizes the total distance required. This problem is known as the traveling salesman problem, and turns out to be very difficult to solve optimally. The calculations required to find the optimal solution grow exponentially with the problem
size (known mathematically as an NP hard problem). In this section, we present a
simple heuristic for obtaining approximate solutions known as the savings method.
6. Risk pooling. One key technique for mitigating uncertainty and improving planning is risk pooling. In essence, this principle states that the sum of a number
of uncertain variables is inherently less variable than each individual variable,
and as such is easier to plan for, schedule, and manage. There are a variety of
ways to operationalize risk pooling, including product postponement, regional
­warehouses, aggregate capacity plans, and flexible capacity.
7. Designing products for supply chain efficiency. “Thinking outside the box” has
become a cliché. It means looking at a problem in a new way, often not taking
constraints at face value. An example of thinking outside the box is postponement
in supply chains. The first application of this idea is due to Benetton, a wellknown manufacturer of fashion knitwear. Benetton must predict consumers’ color
preferences in advance of the selling season. Because wool is dyed first and then
later weaved into sweaters, the color mix must be decided upon well in advance.
If their predictions about consumers’ color preferences are wrong (which they
invariably are), popular colors would sell out quickly and unpopular colors would
sit on the shelves. Their solution was to reverse the order of the weaving and
dying operations. Sweaters were woven from undyed wool (gray stock) and then
dyed to specific colors as late as possible. This provided Benetton with more time
to observe which colors were selling best. Hewlett Packard discovered a similar
solution in their printer division. Printers must be configured for local markets due
to language and other differences. By producing “gray stock” printers that had all
common parts, and then configuring export printers on site in local markets, they
were able to delay product differentiation and better balance their inventories.
Another example of designing products for supply chain efficiency is Ikea. Ikea
is a Swedish-based firm that sells inexpensive home furnishings. To reduce costs
Ikea designs their furniture to be easily stored directly at the retail outlets. This
means that customers can take their purchases with them, thus removing the
long delays and customization required by more traditional furniture outlets.
8. Multilevel distribution systems. Typically in large systems, stock is stored at multiple
locations. Distribution centers (DCs) receive stock from plants and factories and
then ship to either smaller local DCs or directly to stores. Some of the advantages
of employing DCs include economies of scale, tailoring the mix of product to a
particular region or culture, and safety stock reduction via risk pooling.
9. Incentives in the supply chain. Consider a clothing designer whose goods are sold
at a chain of high-end boutiques. The designer contracts the manufacturing to
a firm that subcontracts to a plant in China. The Chinese plant manager is paid
nah77856_ch06.indd 316
17/12/14 6:55 PM
Chapter Six Supply Chain M
­ anagement 317
a bonus based on the quantity of output. As a result, he provides incentives for
his workers to produce as quickly as they can. However, this results in slipshod
quality and a high rate of defective pieces. Ultimately, the designer has to answer
to the boutique chain that carries her designs. This is an example of misaligned
incentives. How could this problem be ameliorated? One possible answer is to
have careful inspection at the plant level, and pay the plant manager based on
nondefective items only. What this means is that each player in the supply chain
needs to have its incentives aligned with what is best for the system as a whole.
10. Global supply chain management. Today, most firms are multinational. Products
are designed for, and shipped to, a wide variety of markets around the world. As
an example, consider the market for automobiles. Fifty years ago, virtually all the
automobiles sold in the United States were produced here. Today, that ­number is
probably closer to 50 percent. Global market forces are shaping the new economy.
Vast markets, such as China, are now emerging, and the major industrial powers
are vying for a share. Technology, cost considerations, and political and macroeconomic forces have driven globalization. Selling in diverse markets presents special
problems for supply chain management.
We take a trip to the grocery store at 10 p.m. to get a jar of peanut butter for our child’s
lunch the next day. Not only is the store open, but there are a large variety of brands,
styles, and sizes of peanut butter available. Americans (and residents of most modern
countries) take such things for granted. However, there is a complex myriad of activities
that must be carefully coordinated to ensure that the peanut butter will be there when
we need it. And this goes for clothes, hardware, and all the other consumer goods we
purchase. Do we appreciate the fact that many of these goods are produced and shipped
all over the world before they make it to our homes? The logistics of coordinating all
the ­activities that afford us this convenience is the essence of supply chain management.
The term supply chain management (SCM) seems to have emerged in the late 1980s
and continues to gain interest at an increasing rate. The trade literature abounds with
book titles and articles relating to some aspect of SCM. Software and consulting firms
specializing in SCM solutions are now commonplace. These companies have grown
at remarkable rates and include giants SAP and Oracle, which offer SCM solutions
as part of comprehensive information retrieval systems. Although the SCM label is
somewhat new, the problems considered are not. Nearly all the material in Chapters
2 to 5 involves SCM. So what is different about SCM? The simple answer is that
SCM looks at the problem of managing the flow of goods as an integrated system.
Many definitions of SCM have been proposed, and it is instructive to examine some
of them. The simplest and most straightforward appears at the Stanford Supply Chain
Forum (1999) Web site and is probably due to Hau Lee, the head of the Forum. It is
Supply chain management deals with the management of materials, information and ­financial
flows in a network consisting of suppliers, manufacturers, distributors, and customers.
While short, this definition is fairly complete. It indicates that it is not only the flow
of goods that is important, but the flow of information and money as well.
A definition due to Simchi-Levi et al. (1999, p. 1) that focuses on only the flow of
goods is
Supply chain management is a set of approaches utilized to efficiently integrate suppliers,
­manufacturers, warehouses, and stores, so that merchandise is produced and distributed
at the right quantities, to the right locations, and at the right time, in order to minimize
­systemwide costs while satisfying service level requirements.
nah77856_ch06.indd 317
17/12/14 6:55 PM
318 Chapter Six Supply Chain Management
Figure 6–1
The supply chain
umbrella
Supply
Operations
Vendor management Production scheduling
Inbound logistics
MRP
Raw materials
In-process inventory control
Capacity planning
Distribution
Forecasting
Customer service
Finished goods
inventory control
Outbound logistics
Since this definition focuses on the flow of goods, it implies that SCM is relevant only
to manufacturing firms. Is SCM also relevant to service organizations? These days it is
well recognized that many aspects of SCM do indeed apply to service organizations, although typically these organizations refer to the value chain instead of the supply chain.
Most writers agree that “logistics” deals with essentially the same issues as supply
chain management. The term logistics originated in the military and concerned problems of moving personnel and materiel in critical times. It was later adopted by business
and became a common moniker for professional societies and academic programs. One
might argue that although classical logistics treated the same set of problems as SCM, it
did not consider the supply chain as an integrated system (Copacino, 1997). A ­summary
of the “umbrella” of activities composing SCM appears in Figure 6–1.
In a sense, one might consider this entire text to deal with some aspect of SCM if
one takes the broad view of the field depicted in Figure 6–1. However, most of the
chapters focus on specific topics, like inventory control or job shop scheduling. SCM,
on the other hand, treats the entire product delivery system from suppliers of raw
­materials to distribution channels of finished goods. While important and useful in
many contexts, simple formulas such as the EOQ are unlikely to shed much light on
­effective management of complex supply chains.
The Supply Chain as a Strategic Weapon
Where does the supply chain fit into the overall business strategy of the firm? In
­Chapter 1, it was noted that an important part of the firm’s competitive edge is its strategic positioning in the marketplace. Examples of strategic positioning include being
the low-cost provider (such as Hyundai automobiles) or the high-quality provider
(such as Mercedes-Benz) or exploiting market segmentation to achieve shares of both
markets (as does General Motors with different brands aimed at different market segments). The design of the supply chain also reflects a firm’s strategic ­positioning.
nah77856_ch06.indd 318
17/12/14 6:55 PM
6.1 Supply Chain Strategy 319
In a supply chain, the primary trade-off is between cost and response time. Ground
transportation (boat, truck, or rail) is less expensive but air freight is faster. Will deliveries be more reliable if the product is moved via the firm’s internal system or would
it be better to subcontract the logistics operation to a third party? Third-party logistics
(abbreviated 3PL) is becoming more common for the same reason that third-party
manufacturing has become so widespread. Companies such as Foxconn, a Taiwanese
multinational electronics manufacturer, are able to achieve significant economies of
scale by providing manufacturing services to large numbers of firms producing similar
products. In this way, it can be less expensive to subcontract manufacturing than to do
it in-house. Cost is not the only issue. For example, Apple Inc. outsources to Foxconn
because they have long realized that their core competencies lie in aesthetics and innovation and not in manufacturing. Cisco Systems Inc., also outsources to Foxconn.
They have transitioned from a manufacturing-focused company to a design-focused
firm. They have also embraced a servicization strategy (see Chapter 1).
Key elements of effective supply chain management include the use of information,
analytics, and incentives. The Snapshot Application featured in this section highlights
Wal-Mart, whose extraordinary success was achieved to some extent because of its
sophisticated supply chain strategies and effective use of information, analytics, and
incentives.
As previously mentioned, one of the key challenges in supply chain management
is aligning incentives. Much of the research is concerned with developing appropriate
contracts to align incentives. Modern technologies make this easier. In fact, it has been
almost 20 years since Gibson, the then chairman of Manugistics Inc., stated in a Wall
Street Journal article that “strategic partnering used to mean stealing revenue or pushing cost onto someone else in the supply chain. You are a pig at the trough if you view
it that way. With technology, there are so many efficiencies that can be shared.” (WSJ,
April 12, 1996)
In this chapter we provide the reader with an appreciation of the most important
issues encountered in managing complex supply chains. In addition, we present a sampling of the mathematical models (analytics) used in SCM analysis.
6.1 Supply Chain Strategy
As just discussed, one of the key trade-offs in supply chain design is between cost and
speed. It is important to recognize that the best choices on this dimension depend on the
product being produced. Fisher (1997) presents a simple framework for this choice. He
breaks products into two categories: functional or innovative, and supply chains into two
types: efficient versus responsive. He argues that functional products should be produced
using efficient supply chains while innovative products require responsive supply chains.
A functional product is a commodity-type product that competes primarily on cost.
A canonical functional product is Campbell’s Soup. Only 5 percent of the products they
produce are new each year and most products have been in the market for many years.
It has highly predictable demand and long life cycles (particularly relative to lead
times). This type of product is best produced using an efficient supply chain where the
primary focus is on costs, which include inventory holding, transportation, handling,
and manufacturing costs. Large manufacturing batches achieve scale economies, and
minimum order quantities reduce handling/order processing costs. Supply chain performance is evaluated using traditional cost measures, such as inventory turns, and
factory and transportation utilization.
nah77856_ch06.indd 319
17/12/14 6:55 PM
Snapshot Application
320 Chapter Six Supply Chain Management
Wal-Mart Wins with Solid Supply
Chain Management
Although there are many examples of companies winning (or losing) because of either good or bad supply
chain strategies, perhaps none is more dramatic than the
stories of Wal-Mart and Kmart. Both companies were
founded in the same year, 1962. In only a few years,
Kmart had become a household term, while Wal-Mart
was largely unknown except for a few communities in
the South. Wal-Mart stores were typically located in rural
areas, so they rarely competed head-on with the larger,
better-known chains like Sears and Kmart, most of
whose stores were located in large cities and surrounding suburbs. In 1987 Kmart’s sales were almost twice
those of Wal-Mart (about $25 billion annually versus
about $15 billion annually). By 1990 Wal-Mart had overtaken Kmart and in 1994 Wal-Mart’s annual sales ­almost
tripled Kmart’s (about $80 billion versus about $27 billion)! Wal-Mart is now the largest discount ­retailer in
the United States, surpassing Sears as well as Kmart, who
have now merged. What could have possibly accounted
for this ­dramatic turn-around?
While one can point to several factors leading to
Wal-Mart’s success, there is no question that the firm’s
emphasis on solid supply chain management was one of
the most important. According to Duff and Ortega (1995)
in comparing the management decisions of Kmart’s former CEO, Joseph Antonini, and Wal-Mart’s Sam Walton:
When Mr. Antonini took the reins of Kmart in 1987, he
had his hands full . . . Also, his predecessors neglected
to implement the sophisticated computer systems
that were helping Wal-Mart track and replenish its
­merchandise swiftly and efficiently.
A self-promoter with a boisterous voice and
wide smile, Mr. Antonini invested heavily in na­tional
television campaigns and glamorous r­ epresentatives
such as Jaclyn Smith . . . Mr. ­Walton avoided publicity.
And instead of marketing, he ­became obsessed with
operations. He invested tens of millions of dollars in a
companywide computer ­system linking cash registers
to headquarters, enabling him to quickly restock goods
selling off the shelves. He also ­invested heavily in trucks
and distribution c­ enters. Besides enhancing his control,
these moves sharply reduced costs.
Mr. Antonini tried bolstering growth by overseeing the
purchase of other types of retailers: the Sports ­Authority
sporting goods chain, ­Office­Max office supply stores, Borders bookstores, and Pace Membership Warehouse clubs . . .
But the least visible difference between Wal-Mart
and Kmart was beginning to matter a lot. Wal-Mart’s
nah77856_ch06.indd 320
incredibly sophisticated distribution ­inventory and
scanner systems meant that ­customers almost never
­encountered depleted shelves or ­price-check delays at
the cash ­register.
The halls of Kmart, meanwhile, were filled with
­distribution horror stories. Joseph R. Thomas, who
oversaw distribution, said that in retrospect, he should
have smelled trouble when he found warehouses
stuffed full of merchandise on ­December 15, the
height of the Christmas ­season.
Wal-Mart is the world’s largest public corporation,
according to the Fortune Global 500 list in 2014, the
biggest private employer in the world with over two
million employees, and the largest retailer in the world.
It has over 11,000 stores in 27 countries, under a total
of 55 different names, and continues to expand. Its
Chinese operations are growing particularly quickly
­
with 100 stores planned for the next few years and a
projected market of 2,000 stores.
Wal-Mart is known for both its sophisticated ­business
analytics and advanced information systems. They have
been a leader in vendor-managed inventory, sharing
data with suppliers for better decisions, implementing
effective cross-docking operations, and lately they have
been proponents of sustainable supply chain management. Although much of the information about their
systems is proprietary, it is clear that they invest heavily
in business intelligence.
There are some important lessons here. When problems arose, Kmart’s management responded by putting
money into marketing and acquisitions, while Wal-Mart
invested in better stores and a better logistics system
for making sure that shelves were stocked with what
customers wanted. One of the key themes of this book
is that operations matter, and indeed effective operations and supply chain management can be a ­strategic
weapon for a firm. While innovative designs can delay
the need for effective supply chain management,
sooner or later inefficiencies catch up with a firm.
The story of Kmart and Wal-Mart is reminiscent of
the story of the American and Japanese auto industries.
In the 1950s and 1960s when American auto makers had
a virtual monopoly on much of the world’s auto market,
where did the Big Three put their r­ e­sources? Into producing better engineered cars at lower cost? No. Into slick
marketing campaigns and annual cosmetic m
­ akeovers.
The Japa­nese, on the other hand, invested in the latest
automotive technology and sophisticated logistics ­systems
such as just-in-time. The rest, as they say, is history.
17/12/14 6:55 PM
6.2 The Role of Information in the Supply Chain 321
An innovative product is one that typically has high margin but highly variable
demand and short life cycle. An example is Sport Obermeyer’s fashion skiwear.
Each year, 95 percent of the products are completely new designs and demand forecasts err by as much as 200 percent. There is a short retail season, and cost concerns
focus around inventory obsolescence and lost sales (see the Snapshot Application
in Chapter 2). This type of product is best produced using a market responsive supply chain, where a premium is paid for flexibility, including faster transportation,
lower transportation and factory utilization, and smaller batches. Supply chain performance is best measured not with traditional cost measures but by recognizing that
the opportunity costs of lost sales and poor service are key metrics to consider.
Cost and speed are both key measures of supply chain effectiveness. However,
Lee (2004) argues that they are not the whole story. He proposes the “Triple A”
supply chain, which is agile, adaptable, and aligned. Here agility is defined as the
ability to respond to short-term changes in demand or supply quickly. Adaptability
is defined as the ability to adjust the supply chain design to accommodate market
changes. Finally, alignment relates to the existence of incentives for supply chain
partners to improve performance of the entire chain. Further details may be found
within his article but clearly these are desirable traits for a supply chain. Like any
desirable feature they are not costless, and the cost to implement must be balanced
against strategic necessity.
6.2 The Role of Information in the Supply Chain
How often have we heard it said that we are living in the “information age”? The availability of information is increasing at an exponential rate. New sources of information
in the form of academic and trade journals, magazines, newsletters, and so on are
introduced every day. The explosion of information availability on the Web has been
truly phenomenal. Web searches are now the first place many go for information on
­almost anything.
Knowledge is power. In supply chains, information is power. It provides the decision maker the power to get ahead of the competition, the power to run a business
smoothly and efficiently, and the power to succeed in an ever more complex environment. Information plays a key role in the management of the supply chain. As we saw
in the earlier chapters of this book, many aspects of operations planning start with the
forecast of sales and build a plan for manufacturing or inventory replenishment from
that forecast. Forecasts, of course, are based on information.
An excellent example of the role of information in supply chains is the Harvard
Business School cases Barilla SpA (A and B) (1994) written by Jan Hammond. In
many operations management courses, these cases provide an introduction to the
role of information in supply chains. Barilla is an Italian company that specializes
in the production of pasta. In the late 1980s, Barilla’s head of logistics tried to
introduce a new approach to dealing with distributors, which he called just-in-time
distribution (JITD). Briefly, the idea was to obtain sales data directly from the
distributors ­(Barilla’s customers) and use these data to allow Barilla to determine
when and how large the deliveries should be. At that time, Barilla was operating
in the traditional fashion. Distributors would independently place weekly orders
based on standard ­reorder point methods. This led to wide swings in the demand
on Barilla’s factories due to the ­“bullwhip effect,” which is discussed in detail
next. The JITD idea met with staunch resistance, both within and outside Barilla.
nah77856_ch06.indd 321
17/12/14 6:55 PM
Snapshot Application
322 Chapter Six Supply Chain Management
Anheuser-Busch Re-engineers Their
­Supply Chain
Anheuser-Busch is the leading American brewer, holding close to a half share of U.S. beer sales, and producing iconic brands such as Budweiser and Bud Light. In
the late 1990s they undertook a major reorganization
of their supply chain (John & Willis, 1998).They found
that less than 10 percent of their volume accounted
for around 80 percent of the brand and package combinations. Management decided to switch to focused
production. Some breweries were dedicated for the
­efficient production of large volume items, such as Bud
and Bud Light. Others were responsible for producing
lower volume niche products, such as wheat beers or
products targeted at holidays. (Refer to the discussion
of the focused factory in Chapter 1.)
By aligning supply chain metrics to the type of product
being produced, their decisions were in line with Fisher’s
recommendations for matching product type with supply chain type. In particular, they made sure that their
6.1 Supply Chain Strategy 322
­ igh-volume breweries were measured by efficiency meth
rics, such as utilization and volume of output; whereas,
their niche product breweries were measured by how
effectively they met demand. Previously, benchmarking
­
had been across all breweries, which did not create the correct incentives for either the high volume or niche products.
In addition to focused facilities and supply chains,
Anheuser-Busch also undertook a number of initiatives
in line with topics covered later in this chapter. In particular, a mixed integer programming model (i.e., supply
chain analytics) was used to identify the optimal number
and location of distribution points in the supply chain.
Contracts with a thousand different trucking companies
(across the whole network) were largely replaced by a
single contract with a single dedicated carrier, decreasing
the complexity of the network. Inventory pooling was
used to reduce both risk and cost. Finally, replenishment
agreements were renegotiated and vendor-managed
­inventory agreements were put in place, which improved
the alignment of the incentives in the supply chain.
The Barilla sales and marketing organizations, in particular, were most threatened by the proposed changes. D
­ istributors were concerned that their prerogatives
would be compromised.
Without going into the details [which can be found in Barilla SpA (B)], management
eventually did prevail, and the JITD system was implemented with several of Barilla’s largest distributors. The results were striking. Delivery reliability improved, and
the variance of orders placed on the factory were substantially reduced. The ­program
proved a win-win for Barilla and its customers.
Barilla’s success with its JITD program is one example of what today is known
as vendor-managed inventory (VMI). VMI programs have been the mainstay of
many successful retailers in the United States. For example, Procter & Gamble
has assumed responsibility for keeping track of inventories for several of its major
clients, such as Wal-Mart. Years ago, grocery store managements were responsible
for keeping track of their inventories. Today, it is commonplace to see, in our local
grocery stores, folks checking shelves who are not store employees but employees
of the manufacturers. With VMI programs, it becomes the manufacturer’s responsibility to keep the shelves stocked. While stock-outs certainly hurt the retailer,
they hurt the manufacturer even more. When a product stocks out, customers typically will substitute another, so the store makes the sale anyway. It is the manufacturer that really suffers the penalty of lost sales, so the manufacturer has a strong
incentive to keep shelves stocked.
The Bullwhip Effect
Barilla’s experience prior to the implementation of their VMI program is one ex­
ample of the bullwhip effect. It has become a topic of considerable interest among
both practitioners and academics alike. The history of bullwhip appears to be the
following.
nah77856_ch06.indd 322
17/12/14 6:55 PM
6.2 The Role of Information in the Supply Chain 323
Executives at Procter & Gamble (P&G) were studying replenishment patterns for one
of their best-selling products: Pampers disposable diapers. They were surprised to see
that the orders placed by distributors had much more variation than sales at r­ etail stores.
Furthermore, orders of materials to suppliers had even greater variability. D
­ emand for diapers is pretty steady, so one would assume that variance would be low in the entire supply
chain. However, this was clearly not the case. P&G coined the term “bullwhip” effect for
this phenomenon. It also has been referred to as the “whiplash” or “whipsaw” effect.
This phenomenon was observed by other firms as well. HP experienced the ­bullwhip
effect in patterns of sales of its printers. Orders placed by a reseller exhibited wider
swings than retail sales, and orders placed by the printer division to the company’s
integrated circuit division had even wider swings. Figure 6–2 shows how ­variance
increases as one moves up the supply chain.
Where does the bullwhip effect come from? One cause stems from a phenomenon
discussed in Chapter 8 in MRP systems. Consider a basic two-level MRP system in
which the pattern of final demand is fixed and constant. Since items are produced
in batches of size Q, the demand pattern one level down is much spikier. The tendency
for lower levels in the supply chain to batch orders in this way is one of the root causes
of the bullwhip effect.
High levels of inventories, caused in part by the bullwhip effect, are common in
the grocery industry as well. To address this problem, the industry has adopted the
­efficient consumer response (ECR) initiative (see, for example, Crawford, 1994). The
total food delivery supply chain, from the point at which products leave the manufacturer to the point when they are stocked on the shelves at the retailer, has an average
of over 100 days of supply. The stated goal of the ECR initiative is to save $30 billion
­annually by streamlining food delivery logistics.
Another example of the bullwhip effect is the popular “beer game” due to S
­ terman
(1989). Participants play the roles of retailers, wholesalers, and manufacturers of beer.
Communication among participants is prohibited: each player must make ordering
Source: H. L. Lee,
P. Padmanabhan, and
S. Whang, 1997.
20
Consumer sales
Order quantity
Increasing ­variability
of orders up the s­ upply
chain
Order quantity
Figure 6–2
15
10
5
0
Time
20
15
10
5
0
nah77856_ch06.indd 323
15
10
5
0
Time
Time
Manufacturer orders to supplier
Order quantity
Order quantity
Wholesaler orders to manufacturer
20
Retailer orders to wholesalers
20
15
10
5
0
Time
17/12/14 6:55 PM
324 Chapter Six Supply Chain Management
d­ ecisions based on what is demanded from the downstream player only. What one
­observes are wild swings in orders placed downstream even when the original ­demands
are fairly stable. This is a consequence of the bullwhip effect.
The telescoping variation in the demand patterns in a supply chain results in a planning nightmare for many industries. What can be done to alleviate these effects? First,
we need to understand the causes of this phenomenon. According to Lee, P
­ admanabhan
and Whang (1997), there are four primary causes of the bullwhip effect:
•
•
•
•
Demand forecast updating
Order batching
Price fluctuations
Shortage gaming
We consider each of these effects in turn. Demand forecasts at each stage of the
supply chain are the result of demands observed one level downstream (as in the
beer game). Only at the final stage of the chain (the retailer) are consumer demands
­observed directly. When each individual in a serial supply chain determines his or
her demand forecasts i­ndividually, the bullwhip effect results. The retailer builds in
safety stocks to protect against uncertainty in consumer demand. These safety stocks
cause the retailers’ orders to have greater variance than the consumers’. The distributor ­observes these swings in the orders of the retailer and builds in even larger safety
stocks, and so on.
Order batching is the phenomenon we saw in MRP systems that results in smooth
demand patterns being translated to spiky demand patterns at lower levels of the product structure. The natural tendency to save fixed costs by ordering less frequently
(which is the idea behind the EOQ formula) gives rise to order batching. Given the
cost structure at each level, this is certainly a reasonable response.
When prices fluctuate, there is a speculative motive for holding inventories. (This
was first discussed in the inventory management context by Arrow, 1958, and ­relates to
the motivations for holding cash postulated by the economist John Maynard Keynes.)
In the food industry, the majority of transactions from the manufacturer to distributors
are made under a “forward buy” arrangement. This refers to the practice of buying in
advance of need because of an attractive price offered by manufacturers. Such practices also contribute to the bullwhip effect. Large orders are placed when promotions
are offered.
Shortage gaming occurs when product is in short supply and manufacturers place
customers on allocation. When customers find out that they may not get all the units
they request, they simply inflate orders to make up for the anticipated shortfall. For
­example, if a computer manufacturer expects to receive only half a request for a CPU
in short supply, they can simply double the size of their order. If anticipated demands
don’t materialize, orders can be canceled. The result is that the manufacturer gets an
­inflated picture of the real demand for the product. This could have dire consequences
for a manufacturer placing large amounts of capital into capacity expansion based on
“phantom” demands.
Clearly, the bullwhip effect is not the result of poor planning or irrational behavior
on the part of players in the supply chain. Each individual acts to optimize his or her
position. What, then, can be done to alleviate the situation? There are several ­poten­tial
remedies, which we will discuss. However, these remedies must take into account
that people behave selfishly. The motivation for bullwhip-inducing behavior must be
­eliminated.
nah77856_ch06.indd 324
17/12/14 6:55 PM
6.2 The Role of Information in the Supply Chain 325
Lee, Padmanabhan, and Whang (1997) recommend four initiatives. They are
1.
2.
3.
4.
Information sharing
Channel alignment
Price stabilization
Discouragement of shortage gaming
1. Information sharing means that all parties involved share information on
point-of-sale (POS) data and base forecasts on these data only. Information sharing can be accomplished by several techniques, one of which is EDI (electronic
data interchange), discussed in detail later. The trend toward information sharing
is beginning to take hold. Computer manufacturers, for example, are now requiring sell-through data from resellers (these are data on the withdrawal of stocks
from the central warehouse). At some point, we expect to see manufacturers linked
­directly to sources of POS data.
2. Channel alignment is the coordination of efforts in the form of pricing, transportation, inventory planning, and ownership between upstream and downstream
sites in the supply chain. One of the tendencies that defeats channel alignment is
order batching. As we noted earlier, fixed costs motivate order batching. Reducing
fixed costs will result in smaller order sizes. One of the important components of
fixed costs is the paperwork required to process an order. With new technologies,
such as EDI, these costs could be reduced substantially. Another factor motivating
large batches is transportation economies of scale. It is cheaper on a per-unit basis
to order a full ­truckload than a partial one. If manufacturers allow customers to
order an assortment of items in a single truckload, the motivation to order large
batches of single items is reduced. Another trend encouraging small batch ordering
is the outsourcing of logistics to third parties. Logistics companies can consolidate loads from multiple suppliers. Outsourcing of logistics (like outsourcing of
­manufacturing) is expanding rapidly.
3. Pricing promotions motivate customers to buy in large batches and store items
for future use; this is very common in the grocery industry. By stabilizing pricing, sales
patterns will have less variation. In retailing, the effect of s­ table pricing is evident by
comparing sales patterns at retailers such as Macy’s that run ­frequent promotions and
warehouse stores such as Costco that offer everyday low pricing. The warehouse stores
experience steadier sales than do the department stores. In fact, for many department
stores, promotional sales account for most of their business. Major grocery manufacturers such as P&G, Kraft, and Pillsbury are moving toward a value-pricing strategy and
away from promotional pricing for the same reason.
4. One way to minimize excessive orders as a result of shortage gaming is to
­allocate based on past sales records rather than on orders. This will reduce the tendency
of customers to exaggerate orders. Several companies (including General Motors) are
moving in this direction.
The bullwhip effect has been observed in a variety of settings and is also theoretically predicted to occur by so-called systems dynamics models (Sterman,
1989). However, a well-designed production system should work to smooth variability, particularly that associated with seasonality. Cachon et al. (2007) perform an empirical study on demand variability across a wide range of U.S. firms.
They find that “industries with seasonality tend to smooth production relative to
nah77856_ch06.indd 325
17/12/14 6:55 PM
326 Chapter Six Supply Chain Management
d­ emand, whereas industries without seasonality tend to amplify.” They also show
that r­etailers, in particular, tend to perform a smoothing rather than amplifying
­effect on demand variability.
The study by Cachon et al. is primarily at the industry level, whereas Bray and
Mendelson (2012) perform a firm level study that finds “65 percent of firms ­exhibit
a positive overall bullwhip.” They confirm the findings of Cachon et al. that in
highly seasonal settings the production smoothing effect outweighs the inherent
­uncertainty amplification caused by demand shocks. Interestingly, they also show
that the bullwhip effect is significantly reduced for data after 1995. While much
of this can be explained by improved information systems, perhaps the increasing awareness of the bullwhip effect is also a contributing factor. Bray and Mendelson (2012) empirically verify the impressive bullwhip reduction achieved by
­Caterpillar Inc., which is known to have focused its efforts on initiatives similar to
the four outlined above, for bullwhip reduction.
In summary, the bullwhip effect is the consequence of individual agents in the
supply chain acting in their own best interests. To mitigate bullwhip effects, several
changes must be made. Incentives must be put in place to reduce demand forecast
­errors, reduce excessive order sizes in allocation situations, and encourage information
sharing and system alignment. As these initiatives become policy, everyone, especially
the consumer, will benefit from reduced costs and improved supply chain efficiency.
Electronic Commerce
Electronic commerce, or simply e-commerce, is a catch-all term for a wide range
of methods for effecting business transactions without the traditional paper-based
systems. It includes EDI, e-mail, electronic funds transfers, electronic publishing,
image processing, electronic bulletin boards, shared databases, and all manner of
Internet-based business systems (Handfield and Nichols, 1999). The point-of-sale
barcode system, now ubiquitous in supermarkets and retailers, is another type of
e-commerce.
Barcodes were first introduced as a way to speed up supermarket checkouts. The
first retail scanning of a barcode appears to be a pack of Wrigley chewing gum in
June 1974 (Varchaver, 2004). However, a powerful advantage of barcodes is their role
in information gathering. Assuming the check-out operator scans the products correctly, the retailer now has accurate information on sales. The reason for the caveat
in the ­previous sentence is that a major source of data error is operators who scan,
for ­example, one can of mushroom soup and hit times three on the till, rather than
separately scanning the mushroom, tomato, and cream of asparagus soups that the
customer has actually purchased. Retailers have put significant educational effort into
ensuring correct scanning techniques by staff. Further, with the introduction of customer loyalty cards, retailers now have information on individual purchasing behavior
and can target promotions to the individual customer.
The internet has fundamentally changed supply chain management. First, it allows
easy transmission of information to users within the firms and to customers and trading partners. This includes point-of-sale demand information, purchase orders, and
­inventory status information. Originally, such information sharing was through electronic
data ­interchange (EDI) systems that allow the transmission of standard business documents in a predetermined format between computers. Dedicated EDI systems have now
been mostly replaced by internet-based systems. Cloud computing has also facilitated
­information sharing and enterprise systems. SAP and Oracle, for example, offer cloud
nah77856_ch06.indd 326
17/12/14 6:55 PM
6.2 The Role of Information in the Supply Chain 327
c­ omputing options for their systems. One of the key advantages of such systems from
a supply chain management perspective is speed. By transmitting information quickly,
lead times are reduced. As noted in Chapter 5, it is the uncertainty of demand over the
replenishment lead time that results in the need to carry safety stock.
In addition to facilitating information sharing, the internet has also provided both
business to consumer (B2C) and business to business (B2B) opportunities. The
rise of B2C commerce was rocky with the now infamous dot-com bust in the early
21st ­century. Many internet-only companies failed at that time; however, Amazon.
com, which is now the world’s largest online retailer, has thrived. Amazon started
as a ­discount bookseller. By selling from a single location, it could reap the benefits
of stock centralization (to be discussed in Section 6.9) and avoid the costly capital
­investment in brick and mortar stores.
The model for internet-based retailing is essentially the same as that for catalog
shopping. Catalog retailers have been around for many years, with Lands’ End and
L.L.Bean being perhaps the most successful. However, catalogs need to be developed,
printed, and mailed regularly, which is quite expensive, especially considering that
most catalogs are thrown away. For this reason, the internet has a significant advantage
over the traditional mail-order catalog business, and both Lands’ End and L.L.Bean
have put significant investment into their websites. Further, the move away from paper
catalogs has allowed them to reach consumers around the world. Lands’ End boasts
that they “proudly ship to over 170 countries.”
Along with the growth of B2C commerce there has also been a less visible growth
of B2B commerce. Hundreds, if not thousands, of firms offer internet-based supply
chain solutions. Many tailor their products to specific industry segments, such as
PrintingForLess.com (printing) and ShowMeTheParts.com (automotive aftermarket).
There are also intermediaries, such as Li & Fung Lmt., who specialize in third party
sourcing. While Li & Fung started as a traditional trading company, today they provide
access to a network of over 15,000 suppliers in more than 40 economies. This would
not be possible without e-commerce.
It is interesting to speculate what effect 3D printing will have on e-commerce.
3D printing, or additive manufacturing, allows the construction of three dimensional
objects by a “printer” just as regular printers allow for the production of two dimensional text. Such printers are currently used widely in the production of manufacturing prototypes but less widely in consumer goods. The range of materials that may
be printed is limited but increasing. Examples of products produced by 3D printing
include custom jewelry and home decor, such as lampshades. In time, instead of buying items from the internet (or store) we may simply buy a design and print it up at
home or in our local print shop. Such a development could drastically simplify supply
chain management!
RFID Technology
Radio frequency identification (RFID) tags are an emerging technology that will
change the way information is stored and transmitted in a supply chain. Unlike barcodes, which are the same across all items of the same stock-keeping unit (SKU),
RFID tags can contain item specific information, such as farm of origin (for produce)
or date of production.
RFID tags were invented in 1973, but are only now becoming commercially ­viable.
They are microchips powered by batteries (active tags) or radio signals (passive tags).
nah77856_ch06.indd 327
17/12/14 6:55 PM
328 Chapter Six Supply Chain Management
The passive device is smaller and cheaper and is likely to emerge as the device of
choice for inventory management applications. Passive tags receive power from the
readers (at which time they are “woken up”) and transmit a unique code. Passive tags
can only be read at close distances but they provide a simple means of electronic identification.
Passive RFID tags cost between $.07 and $.15 each, which makes them too expensive for small grocery items, such as gum, but practical for larger items, such as jeans.
JCPenney announced in 2012 that they were moving to 100 percent RFID tagging of
all items in their stores. While their roll out was postponed, at the time of this writing
they have moved to item-level tagging in bras, footwear, fashion jewelry, and men’s
and women’s denim. Wal-Mart and American Apparel have also moved to tagging
clothing. One of the main advantages of such tagging is in drastically speeding up and
improving accuracy in stock-taking, thereby reducing stock-outs and increasing sales.
For example, in a pilot program at American Apparel Inc. in 2007, sales increased by
14.3 percent (Bustillo, 2010). When all items are tagged retailers will also see savings
in the checkout process because RFID readers can replace the need for item by item
barcode scanning.
Beyond retail item tagging, other applications of RFID technology include (1) EZ
Pass for paying bridge or highway tolls, (2) tagging of luggage on flights, and (3) tagging of cargo containers at most of the world’s ports. As the cost of RFID tags declines,
we will see a much wider range of applications in the context of supply chain management. Reconciling shipments against bills-of-lading (i.e., cargo records) or packing
and pick lists can be performed quickly and accurately electronically, eliminating the
need to perform these functions manually.
Of course, RFID technology has far broader implications than its application to
supply chains. For example, Solusat, the Mexican distributor of the VeriChip—a
rice-sized microchip that is injected beneath the skin—is marketing its device as
an emergency identification. The interest in Mexico for this product is a consequence of the more than 10,000 Mexican children abducted annually. VeriChip
manufacturer, ­Applied Digital Solutions, said it plans to roll out the VeriKid service in other countries, including the United States (Scheeres, 2003). Such applications of RFID technology potentially have enormous benefits, but also can be
used in ways to threaten our privacy.
RFID technology is quickly making inroads into many other industries. Potential
­applications discussed in a recent book on the subject (Schuster, Allen, and Brock,
2007) include the following:
• Warehousing. By tagging inventory stored in a warehouse, improved tracking of
items and order fulfillment can lead to significant improvements in several ­mea­sures
of customer service.
• Maintenance. Maintenance programs require keeping track of the location, s­ uitability,
and condition of spare parts. RFID tags can provide significant ­improvements in monitoring the installed base of parts and components.
• Pharmaceuticals. Tracking and tracing pharmaceuticals can help ameliorate the
problems of counterfeit drugs and theft in the industry. These problems have been
estimated to run to $30 billion annually worldwide.
• Medical devices. RFID technology can provide continuous access to the identity, ­location, and state of medical devices. This has the potential to significantly
­improve patient care.
nah77856_ch06.indd 328
17/12/14 6:55 PM
6.3 The Transportation Problem 329
• Animal tracking. The appearance of mad cow disease (BSE) in the United Kingdom,
Canada, and the United States raised an alarm worldwide about food safety. Being
able to track individual livestock can be invaluable when trying to trace the source
of problems.
• Shelf-life tracking. We are all familiar with expiration dates on foods such as
dairy products, packaged meats, fish, and canned goods. RFID technology provides the opportunity for “dynamic” shelf-life tracking, that is, updating the
shelf-life ­indicator to take into account environmental conditions such as temperature.
• Retailing. It is common for expensive retail items, such as leather jackets, to be
tagged with a transmitter that sounds an alarm if removed from the store. Inexpensive RFID chips provide the opportunity to tag almost all retail items, virtually
eliminating theft altogether.
• Defense. Logistic support has always been a key to securing victory in battle, in
both modern and ancient times. It is currently estimated that the U.S. Department
of Defense manages an inventory valued at $67 billion. Keeping track of this inventory is obviously a top priority, and RFID technology can go a long way toward
accomplishing this goal.
Problems for Sections 6.1 and 6.2
1.
2.
3.
4.
5.
6.
7.
8.
Name two products you are familiar with, that are not discussed in this chapter,
one of which is best classed as functional and the other as innovative.
What product characteristics would be necessary for a product to be able to have
a supply chain that is simultaneously both efficient and responsive?
What is the bullwhip effect? What is the origin of the term?
Do you think that eliminating intermediaries would always get rid of the bullwhip
effect? Under what circumstances might it not work?
Discuss the four initiatives recommended for ameliorating the bullwhip effect in
supply chains.
Amazon.com is a purely e-commerce company, yet Barnes & Noble maintains both retail bookstores and a significant internet business. What are the
advantages and disadvantages of Barnes and Noble’s strategy (as contrasted to
Amazon’s)?
What are the key benefits that the internet has provided in aiding efficiency or
­effectiveness in supply chain management?
Give three examples of revenue or cost benefits made possible by RFID tags when
contrasted to barcodes.
6.3 THE TRANSPORTATION PROBLEM
With good information comes the opportunity to optimize within the supply chain. The
transportation problem is a mathematical model for optimally scheduling the flow of
goods from production facilities to distribution centers. Assume that a fixed amount
of product must be transported from a group of sources (plants) to a group of sinks
(warehouses). The unit cost of transporting from each source to each sink is assumed
nah77856_ch06.indd 329
17/12/14 6:55 PM
nah77856_ch06.qxd
1/10/08
3:45 PM
Page 317
6.1
330 Chapter Six Supply Chain Management
The Transportation Problem 317
to be known. The goal is to find the optimal flow paths and the amounts to be shipped
on those paths to minimize the total cost of all shipments.
The transportation problem can be viewed as a prototype supply chain problem.
Although most real-world problems involving the shipment of goods are more complex, the model provides an illustration of the issues and methods one would encounter
in practice.
Example 6.1
The Pear Tablet
Tablet Corporation produces several types
types of
of tablet
tablet computers.
computers. In 2013,
2013, Pear
Pear
produced
tabletswith
withcapacities
capacities
from
gigabytes
all 10-inch
display.
Thepopular
most
­produced tablets
from
gigabytes
(GB),(GB),
all 10-inch
display.
The most
popular
thetablet.
64 GBPear
tablet.
Pear produces
tabletsplants
in three
plants
in
product product
is the 64isGB
produces
the tabletsthe
in three
located
in S­located
unnyvale,
Sunnyvale,
California;
Dublin,
andThailand.
Bangkok,Periodically,
Thailand. Periodically,
shipments
are
California; Dublin,
Ireland;
andIreland;
Bangkok,
shipments are
made from
made
from production
these threefacilities
production
facilities
to fourwarehouses
distributionlocated
warehouses
in the
these three
to four
distribution
in thelocated
United States
United
States
in Amarillo,
Texas;
Teaneck,
NewIllinois;
Jersey;and
Chicago,
Illinois;
andDakota.
Sioux Falls,
in Amarillo,
Texas;
Teaneck, New
Jersey;
Chicago,
Sioux Falls,
South
Over
South
Dakota.
thebeen
nextdetermined
month, it has
that
thesereceive
warehouses
should
the next
month,Over
it has
thatbeen
thesedetermined
warehouses
should
the following
receive the following
proportions
the company’s
total
production
proportions
of the company’s
totalofproduction
of the
64 GB
tablets. of the 64 GB tablets.
Warehouse
Amarillo
Teaneck
Chicago
Sioux Falls
Percentage of
Total Production
31
30
18
21
The production quantities at the factories in the next month are expected to be (in thousands
of units)
Plant
Sunnyvale
Dublin
Bangkok
Anticipated Production
(in 1,000s of units)
45
120
95
Since the total production at the three plants is 260 units, the amounts shipped to the four
warehouses will be (rounded to the nearest unit)
Warehouse
Amarillo
Teaneck
Chicago
Sioux Falls
Total Shipment
Quantity (1,000s)
80
78
47
55
While the shipping cost may be lower between certain plants and distribution centers, Pear
has established shipping routes between every plant and every warehouse. This is in case of
unforeseen problems such as a forced shutdown at a plant, unanticipated swings in regional
nah77856_ch06.indd 330
17/12/14 6:55 PM
nah77856_ch06.qxd
318
1/10/08
3:45 PM
Page 318
Chapter Six Supply Chain Management
6.3 The Transportation Problem 331
demands, or poor weather along some routes. The unit costs for shipping 1,000 units from
each plant to each warehouse is given in the following table.
Shipping Costs per 1,000 Units in $
TO
F
R
O
M
Sunnyvale
Dublin
Bangkok
Amarillo
Teaneck
Chicago
Sioux Falls
250
1,280
1,550
420
990
1,420
380
1,440
1,660
280
1,520
1,730
The goal is to determine a pattern of shipping that minimizes the total transportation cost
from plants to warehouses. The network representation of Pear’s distribution problem appears
in Figure 6–3.
FIGURE 6–3
Pear Tablet transportation problem
Plants
Sunnyvale
Warehouses
Amarillo
Teaneck
Dublin
Chicago
Bangkok
nah77856_ch06.indd 331
Sioux Falls
17/12/14 6:55 PM
856_ch06.qxd
1/10/08 1/10/08
3:45 PM 3:45
Page PM
321 Page 321
nah77856_ch06.qxd
nah77856_ch06.qxd
1/10/08 3:45
PM Page
321
332 Chapter Six Supply Chain Management
6.2 Solving 6.2
Transportation
Problems withProblems
Linear Programming
321
Solving Transportation
Transportation
with Linear
Linear Programming
Programming
321
6.2 Solving
Problems with
321
Several heuristics for solving transportation problems have been proposed, such
programming.
fact, heuristic
because
theallocates
special
structure
of
thetotransportation
problem,
as
a In
greedy
that
capacity
first
theofcheapest
options. However,
programming.
In fact,
fact,of
because
of the
the
special structure
structure
the transportation
transportation
problem,
programming.
In
because
of
special
of the
problem,
today’s specialized
codes
can
solve
problems
with
millions
of
variables.
The
greedy
it
is
unlikely
that
anyone
with
a
real
problem
would
use
a
heuristic,
since
optimal
today’s
specialized
codes
can
solve
problems
with
millions
of
variables.
The
greedy
specialized
codes can
solve
problems
withamillions
ofhow
variables.
The greedy
heuristic today’s
was
presented
in
Section
6.1
to
give
the
reader
feel
for
flows
are
­heuristic
solutions
can
be
found
efficiently
by ­
l
inear
programming.
In
fact,
because
of are
the
was
presented
in
Section
6.1
to
give
the
reader
a
feel
for
how
flows
heuristic
was
presented
in
Section
6.1reasonable
to give the
reader quickly.
a feel for how flows are
assigned in
networks
and
how
one
could
obtain
solutions
­assigned
special
­structure
of
the
transportation
problem,
today’s
specialized
codes
can
solve
in networks
networks and
and how
how one
one could obtain
obtain reasonable
reasonable solutions
solutions quickly.
quickly.
assigned
in
Let m be
the
number
ofmillions
sourcesof
n the could
number
ofnumber
sinks. (In
Example
6.1,
m 36.1, m 3
problems
with
ofand
variables.
Let
m
be
the
number
sources
and
n
the
of
sinks.
(In
Example
Let m be
the
number of
of the
sources
and nvariables
the number
ofSection
sinks. (In
Example 6.1,
m3
and n 4.)
Recall
the
definition
decision
from
6.1:
Let
m
be
the
number
of
sources
and
n
the
number
of
sinks.
(In
Example
and nn 4.)
4.) Recall
Recall the
the definition
definition of
of the
the decision
decision variables
variables from
from Section
Section 6.1:
6.1:6.1, m 5 3
and
and
4.) from
Recallsource
the definition
the1 decision
variables
from
6.1:
xij n5flow
i tosource
sink of
j for
ij n,1Section
flow
flow from
from
to sink
sink
form11and
ii 1
m
mjand
and
jj n,
n,
xxijij source ii to
j for
1
flow
fromone
source
to sink
1 # i #that
m and
# j cost
# n,of
xij 5
cij as
and defineand
the cost
shipping
unit ifrom
iunit
tojj.for
It follows
the 1total
cij as
define
asofthe
the
cost of
of shipping
shipping
one unit
from
i to
to j.j. ItIt follows
follows
that the
the total
total cost
cost of
of
c
and
define
cost
one
from
i
that
ij
making allmaking
shipments
is
and
define
cshipments
of shipping one unit from i to j. It follows that the total cost
ij as the cost
all
is
making all shipments is
of making all shipments is m n
m nn
m
ij
a a cij xa
accijijxxijij
i1 j1
a
a
i1
j1
i1 j1
For the case
ofthe
Pear
Company,
described described
in Example
6.1,
the objective
function
isfunction is
For
case
of
Pear
Company,
in
Example
6.1,
the
objective
Forthe
thecase
caseof
ofPear
PearCompany,
Company,described
describedin
inExample
Example6.1,
6.1,the
theobjective
objectivefunction
functionisis
For
. . . 1,730x
420x12 420x
380x13 380x
280x14 280x
.
250x11 250x
.
.
.
34
1,730x
.
11
12
13
14
34
250x11
1 420x
420x12
1 380x
380x13
1 280x
280x14
1 .. .. .. 1 1,730x
1,730x34
250x
11 12 13 14 34..
Since many
routes
are
obviously
not
economical,
it
is
likely
that
many
of
the
decision
Since
many
routes
are
obviously
not
economical,
is
likely
that
many
of
the
decision
Since
many
routes
areoptimal
obviously
not economical,
economical, ititit is
is likely
likely that
that many
many of
of the
the decision
decision
Since
many
routes
are
obviously
not
variables will
equal
zero
at the
solution.
variables
will
equal
zero
at
the
optimal
solution.
variables
will
equal
zero
at
the
optimal
solution.
variables
will
equal
zero
at
the
optimal
solution.
The constraints
are designed
to
ensure to
thatensure
the total
amount
shipped
out
of each
The
constraints
are
designed
that
the
total
amount
shipped
out
of
each
The
constraints
are designed
designed
to ensure
ensure
thatamount
the total
total
amount
shipped
out of
of each
each
The
constraints
are
to
that
the
amount
shipped
out
source equals
the
amount
available
at
that
source,
and
the
shipped
into
any
sink
source
equals
the
amount
available
at
that
source,
and
the
amount
shipped
into
any
sink
source
equals
the
amount
available
at
that
source,
and
the
amount
shipped
into
source
equals
the
amount
available
at
that
source,
and
the
amount
shipped
into
any
sink
equals theequals
amountthe
required
atrequired
that sink.
Since
thereSince
are mthere
sources
and
n sinks,
there
are there any
amount
at
that
sink.
are
m
sources
and
n
sinks,
are
sink
the amount
required
at thatSince
sink.constraints).
Since
are ambe
sources
andthere
n sinks,
amount
required
at that
sink.
there
arethere
m sources
and
n sinks,
are
the
total
a total of equals
m
equals
n the
constraints
(excluding
nonnegativity
Let
i
be
the
total
a
total
of
m
n
constraints
(excluding
nonnegativity
constraints).
Let
a
i
there
are
a total
of
m 1 n iconstraints
(excluding
nonnegativity
constraints).
Let
ai be
the total
total
of m
n of
constraints
(excluding
nonnegativity
constraints).
Letsink
ai be
total
toamount
be
shipped
j.into
amount toaamount
be
shipped
source
bj thei and
the
total
to be
beinto
shipped
sink
to
beout
shipped
out of
ofand
source
bjamount
the
total
amount
to
be
shipped
into
the
total
amount
to
be
shipped
out
of
source
i
and
b
the
total
amount
to
shipped
into
sink
j.j.
amount
to
be
shipped
out
source
i
and
b
j
j
Then linear
programming
constraintsconstraints
may be written:
Then
linear
programming
may
be
written:
sink j.linear
Thenprogramming
linear programming
constraints
be written:
Then
constraints
may bemay
written:
n
… m
i xijfor
a xij aa
aa1i … i for
for
… ii …
…m
m
x 11 …
j1
i
a
j1 ij
n
a xij m
n
j1
m
bjm x for
bb1j
a xijij j
a
i1
i1
… j for
… n … j … n
for 11 …
j … n
xij Ú 0 x for
… m
1 …and
j … n. … j … n.
Ú 010 … i for
for
…and
…m
m
xijij Ú
11 …
ii …
and 11 …
j … n.
ForDisk
the Pear
Tablet
Company
problem,
we obtain
the following
seven
constraints:
For the Pear
Drive
Company
problem,
we
obtain
the
following
seven
constraints:
For the
the Pear
Pear Disk
Disk Drive
Drive Company
Company problem,
problem, we
we obtain
obtain the
the following
following seven
seven constraints:
constraints:
For
x11x1
x
1
x
1
x
5
45 (shipments
out
of
Sunnyvale)
12
13
14
x
45
(shipments
out
of
Sunnyvale)
x11 x12 x
13 x12 14 x13 x14 45
(shipments out
out of
of Sunnyvale)
Sunnyvale)
11 xx12 xx13 xx14 45
(shipments
xx11
1
1
5
120 (shipments
out
of
Dublin)
21x1
22
23
24
x
120
(shipments
out
of
Dublin)
x21 x22 x
23 x22 24 x23 x24 120
(shipments out
out of
of Dublin)
Dublin)
21 x
x
x
120
(shipments
xx21
1
x22
x23
1 x24
95 (shipments
out of Bangkok)
32x1
3395
34 5(shipments
out of Bangkok)
x31 x32 x31x
33 x
34 x
xx34
95
95
(shipments
out of
of Bangkok)
Bangkok)
31 32 33 xx32
x
(shipments
out
x31
33
34
1
x2180
1 x31 5(shipments
80 (shipments
into Amarillo)
into
Amarillo)
x11 x21 x11x
31
xx31 80
80
(shipments into
into Amarillo)
Amarillo)
11 xx21
21 (shipments
xx11
x2278
1 x31
78 (shipments
into Teaneck)
12x1
32 5(shipments
into Teaneck)
x12 x22 x
32
x
x
78
(shipments
into
Teaneck)
12 x22
22 x32
32 78
(shipments into
Teaneck)
xx12
x
1
x
5
47 (shipments
into
Chicago)
13x1
23
33
47
(shipments
into
Chicago)
x13 x23 x
33 x23 x33 47
(shipments into
into Chicago)
Chicago)
13 x
x
47
(shipments
xx13
1
x23
1 x33
55 (shipments
into
Sioux Falls)
2455
34 5(shipments
into Siouxinto
Falls)
x14 x24 x14x
34 x
xx34
55
(shipments
Sioux Falls)
Falls)
14 24 x24
(shipments
into Sioux
x14
34 55
and the nonnegativity
constraints
required
in linear programming:
and the nonnegativity
constraints
required
in
linear
programming:
and the
the nonnegativity
nonnegativity constraints
constraints required
required in
in linear
linear programming:
programming:
and
x
$
0 for
1
#
i
#
3 and 1
# j # 4.
ij
and
j 4.1 xij 0 x for
010 i for311 313 and
and
4.
4.
for
ii 1 jj xijij The
problem
was
entered
in
Excel
Solver.
The
spreadsheet
used
and the solution
The problem
entered
Excel Solver.
spreadsheet
used and used
the solution
The was
problem
wasinentered
entered
in Excel
ExcelThe
Solver.
The spreadsheet
spreadsheet
and the
the solution
solution
The
problem
was
in
Solver.
The
used
and
a
­
ppear
in
Figure
6–4.
The
solution
obtained
is
appear in Figure
6–3.
The solution
obtained
isobtained is
appear
in
Figure
6–3.
The
solution
appear in Figure x6–3.5The
solution 5
obtained
is 5 78, x 5 38,
45, x
42, x
x14 45, x 14 x2145,
42, x 21 x2242,
78, x 22 x 78,
38, 31 38,
14 45,
21
22 31
31 38,
x14
x
42,
x22
78,10, xx31
215 47, and x
5
x
x34 10, x 34 10,
x33 47, x 33and
47,
and
33 47, and x34
34 10,
with all other values equalingx33
zero. The total cost of
this solution is $297,800.
i1
nah77856_ch06.indd 332
17/12/14 6:55 PM
nah77856_ch06.qxd
1/10/08
3:45 PM
Page 322
nah77856_ch06.qxd
1/10/08
3:45 PM
Page 322
322 Chapter Six Supply Chain Management
6.4 Generalizations of the Transportation Problem 333
322 Chapter Six Supply Chain Management
FIGURE 6–4
Solution of Pear’s transportation problem using Excel Solver
FIGURE 6–4
A
B
C
D
E
F
G
H
Solution of Pear’s
transportation
problem
using
Excel
Solver
1
2
1
3
2
4
3
5
4
6
5
7
6
8
7
9
8
10
9
11
10
12
11
13
12
14
13
15
14
16
15
17
16
18
17
19
18
20
19
20
O
P
Variables
I
J
K
L
M
N
Solution of Example 6.1 (Pear’s Transportation Problem)
B
C
D
E
F
G
H
I
J
K
L
M
N
of Example
6.1 23
(Pear’s24
Transportation
11 12 Solution
13 14
21 22
31 32 Problem)
33 34 Operator
Value
RHS
Variables
Values
11
0 12
0 13
0 14
45
23
0 24
0 31
0 33
38 32
47 34
10 Operator
Value
RHS
Values
Objective Coeff
st
Objective Coeff
Constraint
1
st
Constraint 2
0
250
0
420
0
380
45 1280
42
280
0 1520
0 1550
0 1660
78 1440
38 1420
47 1730
10
990
Min
297800
250
1
420
1
380
1
280
1 1280
1
1
1
1
990 1440 1520 1550 1420 1660 1730
Min
297800
45
120
45
95
120
80
95
78
80
47
78
55
47
55
A
Constraint
Constraint
Constraint
Constraint
Constraint
Constraint
Constraint
Constraint
Constraint
Constraint
Constraint
Constraint
1
3
2
4
3
5
4
6
5
7
6
7
1
1
1
1
1
1
1
1
1
1
1
1
Notes:
Notes:
21
42 22
78
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
O
P
45
120
45
95
120
80
95
78
80
47
78
55
47
55
Formula for Cell O9: SUMPRODUCT (B9:M9,$B$5:$M$5). Copied to O10 to O15.
Changing cells for Solver are $B$5:$M$5.
Formula for Cell O9: SUMPRODUCT (B9:M9,$B$5:$M$5). Copied to O10 to O15.
Changing cells for Solver are $B$5:$M$5.
with all other values equaling zero. The total cost of this solution is $297,800. Notice
that this is not the same as the solution obtained from the greedy heuristic. As noted
with
allthe
other
zero. The
cost of
this solution
is $297,800.
Notice
earlier,
costvalues
of theequaling
greedy heuristic
is total
$304,900,
which
differs from
the optimal
cost
that
this
is
not
the
same
as
the
solution
obtained
from
the
greedy
heuristic.
As
noted
by only 2.4 percent.
earlier, the cost of the greedy heuristic is $304,900, which differs from the optimal cost
6.4 Generalizations
of the Transportation Problem
by only 2.4 percent.
6.3
6.3
GENERALIZATIONS OF THE TRANSPORTATION PROBLEM
The Pear Company example is the simplest type of transportation problem. Every link
GENERALIZATIONS
OF example
THE
TRANSPORTATION
PROBLEM
The Pear
Company
is and
the simplest
of transportation
Every
link
from
source
to sink is
feasible,
the total type
amount
available fromproblem.
the sources
exactly
from source
to sink
is feasible,
the total
amount
available
from the can
sources
exactly
equals
the total
demand
at theand
sinks.
Several
of these
requirements
be relaxed
The
Pear
Company
example
is the
simplest
typeofofthese
transportation
problem.
Every
link
equals
the
total
demand
at
the
sinks.
Several
requirements
can
be
relaxed
­without making the problem significantly more complicated.
from
source
to
sink
is
feasible,
and
the
total
amount
available
from
the
sources
exactly
without making the problem significantly more complicated.
equals the total demand at the sinks. Several of these requirements can be relaxed
without making the problem significantly more complicated.
nah77856_ch06.indd 333
17/12/14 6:55 PM
334 Chapter Six Supply Chain Management
Infeasible Routes
Suppose in the example that the firm has decided to eliminate the routes from
Dublin to Chicago and from Bangkok to Sioux Falls. This would be accounted
for by placing very high costs on these arcs in the network. Traditionally,
­uppercase M has been used to signify a very high cost. In practice, of course, one
would have to assign a number to these locations. As long as that number is much
larger than the other costs, an optimal solution will never assign flow to these
routes. For the example, suppose we assign costs of $1,000,000 to each of these
routes and re-solve the problem. The reader can check that one now obtains the
following solution:
x14 5 45, x21 5 32, x22 5 78, x31 5 48, and x33 5 47,
with all other values equaling zero. The cost of the new solution is $298,400, only
slightly larger than the cost obtained when all the routes were feasible.
Unbalanced Problems
An unbalanced transportation problem is one in which the total amount shipped
from the sources is not equal to the total amount required at the sinks. This can arise
if the demand exceeds the supply or vice versa. There are two ways of h­ andling
unbalanced problems. One is to add either a dummy row or a dummy column to
absorb the excess supply or demand. A second method for solving unbalanced
problems is to alter the a­ ppropriate set of constraints to either # or $ form. Both
methods will be illustrated.
Suppose in Example 6.1 that the demand for the tablets was higher than anticipated. Suppose that the respective requirements at the four warehouses are now
­Amarillo, 90; Teaneck, 78; Chicago, 55; and Sioux Falls, 55. This means that the
total ­demand is 278 and the total supply is 260. To turn this into a balanced problem, we add an additional fictitious factory to account for the 18-unit shortfall.
This can be labeled as a dummy row in the transportation tableau and all entries
for that row assigned an arbitrarily large unit cost. Note that when supply e­ xceeds
demand and one adds a dummy column, the costs in the dummy column do not
have to be very large numbers, but they do all have to be the same. (In fact, one can
assign zero to all costs in the dummy column.) In the example, we assigned a cost
of 10 6 to each cell in the dummy row. The resulting Excel spreadsheet and Solver
solution (shown in the row labeled “Values”) appear in Figure 6–5. The optimal
solution calls for assigning the shortfall to two warehouses: 8 units to Chicago and
10 units to Sioux Falls.
Unbalanced transportation problems also can be formulated as linear programs by
using inequality constraints. In the previous example, where there is excess demand,
one would use equality for the first three constraints, to be sure that all the supply is
shipped, and less than or equal to constraints for the last four, with the slack accounting for the shortfall. The reader should check that one obtains the same solution by
doing so as was obtained by adding a dummy row. This method has the advantage
of giving an accurate value for the objective function. (In the case where the supply
­exceeds the demand, the principle is the same, but the details differ. The first three
­supply constraints are converted to greater than or equal to form, while the last four
­demand constraints are still equality constraints. The slack in the first three constraints
corresponds to the excess supply.)
nah77856_ch06.indd 334
17/12/14 6:55 PM
nah77856_ch06.qxd
324
1/10/08
3:45 PM
Page 324
Chapter Six Supply Chain Management
6.5 More General Network Formulations 335
FIGURE 6–5
Solution of Example 6.1 with excess demand and dummy row
Solution of Example 6.1 with Excess Demand and a Dummy Row
Variables
Values
Obj Func
11 12 13 14 21 22 23 24 31 32 33 34 41
0
0
0
45
42
78
0
0
48
0
47
0
42
0
43
0
Oper Value
44
8
RHS
10
250 420 380 280 1280 990 1440 1520 1550 1420 1660 1730 1.E06 1.E06 1.E06 1.E06 Min 2E07
st
Constraint 1
1
1
1
1
Constraint 2
1
1
1
1
Constraint 3
1
1
1
1
Constraint 4
Constraint 5
1
1
1
Constraint 6
Constraint 7
Constraint 8
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
45
60
120
130
95
95
18
18
90
90
78
78
55
55
55
55
6.5 MORE GENERAL NETWORK FORMULATIONS
The transportation problem is a special type of network where all nodes are either supply nodes (also called sources) or demand nodes (also called sinks). Linear programming also can be used to solve more complex network distribution problems as well.
One example is the transshipment problem. In this case, one or more of the nodes in the
network are transshipment points rather than supply or demand points. Note that a
transshipment node also can be either a supply or a demand node as well (but no node
is both a supply and a demand node).
For general network flow problems, we use the following balance of flow rules:
If
1. Total supply total demand
2. Total supply total demand
3. Total supply total demand
Apply the Following Rule at Each Node:
Inflow outflow supply or demand
Inflow outflow supply or demand
Inflow outflow supply or demand
The decision variables are defined in the same way as with the simple transportation
problem. That is, xij represents the total flow from node i to node j. For general network
flow problems, we represent the supply as a negative number attached to that node and
the demand as a positive number attached to that node. This convention along with the
flow rules will result in the correct balance-of-flow equations.
Example 6.2
nah77856_ch06.indd 335
Consider the example of Pear Tablets. The company has decided to place a warehouse in
Sacramento to be used as a transshipment node and has expanded the Chicago facility to also
allow for transshipments. Suppose that in addition to being transshipment nodes, both
Chicago and Sacramento are also demand nodes. The new network is pictured in Figure 6–6.
Note that several of the old routes have been eliminated in the new configuration.
17/12/14 6:55 PM
nah77856_ch06.qxd
1/10/08
3:45 PM
Page 325
6.4
336 Chapter Six Supply Chain Management
More General Network Formulations 325
FIGURE 6–6
Pear Tablets problem with transshipment nodes
Plants
Sunnyvale 1
45
Transshipment
nodes
250
76
38
0
Dublin 2
130
25
4
Sacramento
1440
1660
47
5
Chicago
Demand
nodes
6 Amarillo
80
110
0
18
7 Teaneck
78
120
195
95
Bangkok 3
95
8 Sioux Falls
55
We define a decision variable for each arc in the network. In this case, there are a total of
10 decision variables. The objective function is
Minimize
250x16 � 76x14 � 380x15 � 1,440x25 � 1,660x35
� 110x46 � 95x48 � 180x56 � 120x57 � 195x58
The total supply available is still 260 units, but the demand is 285 units (due to the additional 25 units demanded at Sacramento). Hence, this corresponds to case 2 of the flow rules
in which total demand exceeds total supply. Applying rule 2 to each node gives the following
eight constraints for this problem:
Node 1:
�x14 � x15 � x16 � �45
Node 2:
�x25 � �120
Node 3:
�x35 � �95
Node 4:
x14 � x46 � x48 � 25
Node 5:
x16 � x46 � x56 � x56 � x57 � x58 � 47
Node 6:
x16 � x46 � x56 � 80
Node 7:
x57 � 78
Node 8:
x48 � x58 � 55.
For some linear programming codes, all right-hand sides would have to be nonnegative. In
those cases, one would multiply the first three constraints by �1. However, this is not required
in Excel, so we can enter the constraints just as they appear.
The Excel spreadsheet and solution for Pear’s transshipment problem appears in Figure 6–7.
Note that the solution calls for shipping all units from the sources. Since this problem had more
supply than demand, it is interesting to see where the shortfall occurs. The amount shipped into
Sacramento is 45 units (all from Sunnyvale), and the total shipped out of Sacramento is 20 units.
Thus, all the demand (25 units) is satisfied in Sacramento. For the other transshipment point,
Chicago, the shipments into Chicago are 120 units from Dublin and 95 units from Bangkok, and
nah77856_ch06.indd 336
17/12/14 6:55 PM
nah77856_ch06.qxd
1/10/08
4:10 PM
Page 326
326 Chapter Six Supply Chain Management
6.5 More General Network Formulations 337
FIGURE 6–7
Excel spreadsheet for Pear transshipment problem in Example 6.2
A
B
C
1
D
E
F
G
H
I
J
K
L
M
N
Operator
Value
RHS
Min
361530
Excel Spreadsheet for Pear Transshipment Problem in Example 6.2
2
3
Variables
14
15
16
25
35
46
48
56
57
58
Values
45
0
0
120
95
0
20
80
78
10
7
Obj Funct
76
380
250
1440
1660
110
95
180
120
195
8
st
1
1
1
4
5
6
9
Node 1
10
Node 2
11
Node 3
12
Node 4
13
Node 5
14
Node 6
15
Node 7
16
Node 8
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
45
Download